Baltor
Get started
For engineers, designers and AI agents
Give each step research, ready code and tools from one searchable library. Build with files you can inspect and reuse, instead of rewriting the same work.
Start with research: try Agent Feeds, free through 2026
Choose your plan
01 · Research
Agent Feeds
Choose tools and models with sources you can check.
$9 a monthstandard price
Free through December 31, 2026 (Eastern)
Curated source collections in JSON Feed, Markdown and RSS for each agent.
Explore Agent FeedsNo automatic charge. A paid subscription needs your opt-in.
02 · Build
Give each step the files it needs, ready to use.
$49 a month for the whole library
Agent Feeds + Harness Files
The whole library: skills, tools, code and settings. Exact versions through MCP or the API.
Get startedModel usage is separate. Cancel any time from your account page.
03 · Keep working Public preview
Prepare a supervised queue for a worker you control and review its results.
$99 a monthMax plan
Agent Feeds + Harness Files + Overnight / AFK Work
Local overnight tools in Public preview: call and time limits, task checkpoints and morning reports.
Try the free public previewThis preview is free. The local tools remain included with Pro at no extra charge. Max subscriptions are not sold during this public preview.
Engineers
Find scripts, tools and tested methods for software, data and infrastructure.
Explore engineeringDesigners
Explore procedural tools and structured creative briefs.
Playable creative demo
Explore creative workAI Agents
Search for what a step needs and download exact files, or read source-backed feeds with no key.
Connect your agent# search
intelligence_search query "plan an API integration and write its contracts"
# download the exact version
provisioning_read
Packaged bytes match · not a live request
# place in step-1-plan-api-integration/
Example layout
step-1-plan-api-integration/ ├── CLAUDE.md ├── .claude/skills/ │ └── orient-on-a-task-and-write-its-contracts/ │ └── SKILL.md ├── .mcp.json └── .baltor/step.lock.json
The Baltor Harness is built to assemble a folder like this for each step.
Example objective: define the API integration's inputs, permissions and acceptance checks. Work this step addresses: rewriting the plan and its handoff rules for every integration.
Start with a setup guide
Connect your harness, choose useful components and add them to your project. Find your setup steps
Search the library and retrieve selected components through the Model Context Protocol. Check your tool's connection guide
See it work
Open an editable asset, connect an agent, or prepare the research for a release.
Creative production
Combine geometry, materials and lighting. Explore native renders, inspect their source files, and open an editable scene in your own tools.
Explore the creative galleryAgent integration
Connect through MCP or OpenAPI, inspect the contract, and make a checked first download with the same permissions across tools.
Explore the connection workflowDecisions that stay current
Compare models, tools and creative engines against sources your agent can read. Keep the criteria and the trigger for the next review.
Build an agent reading profileMeasured in the open. A real model and harness run · Quality and token comparison · A complete competition workflow
1,560,856files in the library today
Identical shared files are counted once.
Skills with their scripts, instruction files, tools and code, subagents and commands, hooks and MCP server settings, searched from inside your harness. Each download includes its source, licence and exact version.
SKILL.md · scripts/ · assets/
AGENTS.md · CLAUDE.md
scripts, functions and utilities
agent and command definitions
checks your harness runs itself
the tools a step may call
Your harness calls your models through the provider connection you control.
Every download identifies the selected version and includes a checksum to verify its files.
Your harness runs selected components with the file, command and network permissions you choose.
Anything else? Ask on the support page.
Start with the guides for Claude Code, Codex, OpenCode, Pi and Baltor Harness. The setup page also has ready entries for Cursor, VS Code and Gemini CLI, and any other app that speaks MCP connects to the same address.
Three ways. Connect your harness over MCP to search and download files, call the REST API with a Baltor key from a script, or give an agent a public feed link that needs no key. Get set up shows each one.
No. Bring the model access you already have: a provider key, a model on your own computer, Ollama Cloud, or a machine on your network. Library access needs no model key. If you choose the cloud model proxy, you supply a provider key with each request.
One item at the version you chose counts once a month. A search, a second look at the same result and a refused download are not downloads.
Baltor records account and download information, plus feedback, reports and files you choose to submit through available features. Browsing and downloading do not automatically upload your local project or model keys. The privacy notice explains storage, access, data exchange and retention.
Search the whole library from the harness you already use.
Savings depend on the task, the model and the setup.
Use cases
Find reusable files for your project, keep the source, and work in the tools you choose.
Bring reusable code, checks and technical context into your existing project.
Find validation methods, inspect their assumptions and compare the result with your own data checks.
Read the measured comparisonCompare documented models and MCP servers, then use the setup that fits your harness and runtime.
Explore the MCP directoryUse the downloadable worker with your own model access and a record of each step. Library selection remains a separate setup step.
Set up the workerExplore editable prototypes, a playable demo and reusable methods for creative briefs.
Combine scene generation, camera controls and geometry checks, while keeping the editable source.
See the creative workflowPlan reusable camera moves, timing and format changes for product clips and explainers. Export behavior needs testing in the chosen renderer.
Explore motion componentsUse typed parameters and checks for spacing, typography and aspect ratios instead of rebuilding every variation.
Browse available filesDownload only the files a step needs, with explicit versions and permissions.
Search the library for existing methods and implementations, inspect the references and load only the selected files.
Read the retrieval guideDownload native files with their manifests, dependencies and digests. The harness retains control over whether and how to execute them.
Connect an agentUse a focused check to decide the next step, preserve its record and keep any additional work within the task's budget.
See how run records workMCP shortlist
Selected integrations for coding, browsing, research, 3D work and automation. Each has a clear job, public source and setup documentation.
Ordered by GitHub stars within this shortlist. Popularity is not a quality score, an active-user count or proof of use by Baltor.
Public repository signals checked . Stars rank this shortlist only.
Find current library documentation
Retrieve documentation and examples for the library a coding task actually uses.
Publisher-maintained · Checked 2026-10-20
Check the selected library and version. Documentation retrieval does not test generated code.
Inspect and operate a browser
Give a coding harness browser interactions for debugging and checking a web experience.
Publisher-maintained · Checked 2026-10-20
Browser access can expose signed-in pages and trigger changes. Use a scoped browser profile.
Work with repositories and issues
Let a harness inspect repository context and use supported GitHub operations.
Publisher-maintained · Checked 2026-10-20
Enable only the toolsets and account permissions needed for the task.
Inspect and edit a Blender scene
A community bridge for object, material and scene operations inside Blender.
Community integration · Checked 2026-10-20
It can execute Python in Blender. Keep the connection local and test changes on a copy of the project.
Build and validate n8n workflows
A community integration that supplies n8n node information and workflow-building operations.
Community integration · Checked 2026-10-20
This is not n8n's own server. Workflow management depends on the connected instance and its credentials.
Search and extract web content
Use supported search and extraction operations to gather information for a research task.
Publisher-maintained · Checked 2026-10-20
Service access and usage charges are separate from Baltor. Source access and reuse restrictions still apply.
These are editorial starting points for different tasks, with repository-maintained setup instructions and a dated public popularity signal. No paid placement affects inclusion or order. Test the chosen server against your task, permissions and installed versions.
We do not yet publish per-server Baltor usage counts or an adoption ranking. Listing a server here does not mean Baltor agents have used or qualified it. Recorded integrations and successful task runs are required before we label one “used by Baltor.”
The MCP reference-server collection includes examples such as filesystem access and memory. Its repository stars cover the collection, so they are not compared with individual servers above.
Run a local queue of tasks with your own model access. Set a call budget and time limit, keep task checkpoints and review a morning report.
Example morning report07:00
Set model-call and time limits, preserve task checkpoints and review morning reports. Review interrupted attempts before restarting them. The local harness is free.
Read the overnight setup guideMax · Public preview
Prepare a bounded queue for your own Linux, macOS or WSL worker. The same command works on a cloud machine you control. Download the task list, run the command there, and bring its status report back here.
Preparation and report inspection happen in this browser. Your task paths and imported reports stay here. The worker uses your model access and hosting. A queue limit counts calls reserved, so review your provider's token and spending controls separately.
Choose task files and limits for the worker you control.
Save tasks.txt in your checkout. Configure your model access and follow the worker prerequisites and recovery guide before starting. Use Ctrl+C in that worker's terminal to interrupt its owned process group.
Save the command's JSON output as a status file. Reuse the same task paths and limits. The worker preserves reservations and holds uncertain outcomes for reconciliation.
Choose a queue status report to inspect its attempts.
| Task | State | Calls reserved | Process exit |
|---|
See which attempts finished, failed or need review, with their retained task checkpoints.
A finished process is not an accepted result. Check the artifacts against the task's acceptance criteria.
Use your provider account or local endpoint. Call limits control how many requests the queue can make.
Turn it on with a qualified fresh-instance profile, or leave it off for a session-based setup. Check your selected harness's settings. Start with one short task and inspect its artifacts before leaving a longer queue unattended.
A fresh process separates context; it does not establish task success. The worker guide explains the supported profiles and their limits.
Pro, $49 a month. The Baltor Harness is free.
Each step carries only the context, files and tools it asks for. Measure the effect on a task you know: savings depend on the task, the model and the setup.
Context
Search returns short results. Your agent downloads only the item it chose, so a small step never carries the whole history.
Model
You choose the model for each step, so a smaller or local model can take the steps it handles well, and a large model is kept for the steps that need it.
Code
Every item carries its source, licence and exact version, so a script or tool that already exists can be downloaded instead of written again at every run.
Tools
A step sees only the tools and MCP servers it asked for, so its instructions stay short and its choices stay clear.
Savings depend on the task, the model and the setup. We claim no percentage.
In one measured test, library items did not make a cheap model better at data cleanup, cost 45 to 446 percent more prompt tokens per step, and one made results worse. We published it anyway.
Read the data cleanup case studyPro, $49 a month. No overage billing.
Each run on the Baltor Harness records what a step asked for, which files it used, which model ran it, how long it took and whether the result was accepted. Baltor is designed to use these records to set up the next task better.
Record
What the step asked for, which files it used, which model ran it, how long it took and whether the result was accepted.
Next task
Baltor is designed to use these records to choose the files a step is offered, the model it runs on and how the task is broken down. Automatic setup of every step is not available yet.
Evidence
When we compare two ways of doing a step, we decide on recorded results, and we publish a result that says no.
Library
New skills, tools and plugins join the library, and problem reports from customers withdraw broken items.
Pro, $49 a month. Cancel any time.
One downloadable container includes OpenCode, the Baltor harness, Python, Node.js and Git. Your project and model access stay with you.
Install Docker, then pull the worker and check its installation. No model or provider key is included.
docker pull ghcr.io/alisonjieli-png/loop-engine:main docker run --rm --read-only --cap-drop ALL --security-opt no-new-privileges \ --tmpfs /tmp:rw,nosuid,size=512m \ ghcr.io/alisonjieli-png/loop-engine:main doctor --format json
Download the Compose configuration. Save it beside a new project folder. On Linux, set BALTOR_UID and BALTOR_GID to your user and group IDs before mounting your own files. The current published image is Linux amd64; Docker Desktop can use that platform on other supported computers.
Put your instructions in project/task.md. Supply your provider key through your local environment. This example selects an Ollama Cloud model and allocates thirty model calls.
docker compose run --rm --entrypoint baltor-worker worker \ --task /work/task.md --provider ollama_cloud --model gemma4:31b \ --harness baltor --max-model-calls 30 --max-passes 4 \ --authorize-project-commands
Choose --harness opencode for OpenCode, or --harness gateway for the direct provider route. Each run gets a new folder under project/baltor-runs, with its artifacts and saved history. Project commands require the explicit flag above; use it for work you trust inside the mounted project.
The container profiles start a fresh process, home and configuration for each modelled step. Provider credentials stay in the parent broker. The engine can plan additional work and dependency-bound subproblems as a task proceeds, within the same remaining authority and call allowance.
Docker isolates the whole worker. These profiles do not promise a separate operating-system sandbox for every process. Native harness tools are disabled in the brokered profile; the engine owns actions and acceptance. The existing namespace profile remains a separate option on a compatible host.
Read the worker operating guide. Connecting to the hosted library uses a separate personal Baltor client key; downloading this worker does not include a library subscription or model usage.
Try procedural 3D examples and build a reusable creative brief. Adjust a scene, keep its parameters and export an editable result.
The examples below are editable prototypes. The playable browser demo lets you change a scene and export it.
These prototypes generate a static 3D blockout and four reference views from width, height and depth, ready to take into your own rigging, material or simulation tools.
Keep character shape, materials, palette, expression, pose, camera, lighting and output format in a structured brief. Use it to describe revisions consistently.
Use the AI and editing tools you already work with. Baltor supplies selected files; a library subscription does not include model usage, rendering or third-party software licences.
Browse creative filesBrowse the librarySee supported connections
For a developer whose agent begins every task empty-handed. Each step gets the skills, instructions and reusable code it needs, and you keep your own harness, your own models and your own machine.
Your agent searches the library and downloads only the item it chose, keeping each prompt focused on the current step.
Every item carries its source, version, licence and digest, so your agent can pick one up instead of writing it a second time.
The model choice stays yours at every step. Your model key stays in your environment, and Baltor never asks for it.
Your harness runs where it already runs. Baltor answers requests for selected files while your code and repository stay on your machine.
Claude Code, Codex, OpenCode, Pi and the Baltor Harness each have written steps on Get set up.
Pro is $49 a month for the whole library. One downloaded item is the measured unit, with no overage billing.
Pro, $49 a month. No overage billing.
For a team lead who wants the same files in front of every person and every agent, whatever harness each one runs.
Skills, instructions, tools and code sit in one library, each with its source, version, licence and digest attached.
Automated checks of licence, source, safety and declared effects run before a file reaches your agents. How checks work
Each person creates keys for the clients they connect and can revoke one at any time. Revoking takes effect on the next request.
Each person on the team holds their own account and keys, for $49 a month.
Then share Get set up with your team.
Baltor is a hosted library of files for the tools that write your software: skills, instructions, tools and code, each placed where your harness reads it. Your own harness and your own models do the work.
Pro is $49 a month. One downloaded item is the measured unit, with no overage billing.
Agent Feeds alone is $9 a month and free through December 31, 2026 (Eastern), with no automatic charge. The local overnight tools are a free Preview.
One item at the version you chose counts once a month. A search, a second look at the same result and a refused download are not downloads.
For each item your agent downloads, your account keeps one usage record: the item's name, its digest and the time.
One downloaded item is the unit. No overage billing.
Baltor is one authenticated endpoint that serves files directly to the tools and frameworks you already run.
https://baltor.ai/mcpBALTOR_SERVICE_TOKEN in your configurationA search returns short results and their facts, never file contents. A file arrives only through a separate authorized download that names the SHA-256 of the exact bytes.
A request that names a protocol version this service does not serve is answered with the versions it does serve, so your client chooses one. Nothing is changed for you in silence.
Before any setup, the capabilities record names the versions, the transport, the search modes and the download limits.
curl https://baltor.ai/api/v1/capabilities
Serving and connectionsSearching and retrievingEndpoints and local runtimes
Pro, $49 a month. A Baltor key for each tool.
A data cleanup task, broken into five small steps. Each step searches the library, downloads only the file it chose and puts it where the harness reads it.
Recorded starter-catalogue results, not a live search. The current library can return different versions. Always use the digest returned by a fresh search before downloading. About these recorded results.
The task
Clean the customer file customers.csv: learn what its columns hold, split the addresses, find customers entered twice, propose merges without deleting a row, and hand over a checked copy.
Show the folders for
Step 1 of 5
Count what each column holds before anything changes.
Recorded starter-catalogue results
Search
search: profile a text column before cleaning
Download
download profile_
Placed where the harness reads it
Example layout
step-1-profile/ ├── customers.csv the task's file └── .claude/skills/ └── profile-text-column-before-cleaning/ └── SKILL.md the skill this step chose
Writes profile.json for step 2.
Step 2 of 5
Break each address line into number, street, city, region and postal code.
Recorded starter-catalogue results
Search
search: split address lines in a customer file
Download
download split_
Placed where the harness reads it
Example layout
step-2-split/ ├── customers.csv the task's file ├── profile.json from step 1 └── .claude/skills/ └── split-address-lines-into-components/ └── SKILL.md the skill this step chose
Writes addresses.csv for step 3.
Step 3 of 5
Find the customers entered more than once.
Recorded starter-catalogue results
Search
search: find duplicate customer records
Download
download find_
Placed where the harness reads it
Example layout
step-3-duplicates/ ├── customers.csv the task's file ├── addresses.csv from step 2 └── .claude/skills/ └── find-duplicate-records-with-blocking-keys/ └── SKILL.md the skill this step chose
Writes duplicate-pairs.csv for step 4.
Step 4 of 5
Propose one survivor for each group of duplicates, without deleting a row.
Recorded starter-catalogue results
Search
search: propose a dedupe without deleting rows
Download
download propose_
Placed where the harness reads it
Example layout
step-4-merge/ ├── duplicate-pairs.csv from step 3 └── .claude/skills/ └── propose-a-dedupe-without-deleting-rows/ └── SKILL.md the skill this step chose
Writes merge-proposal.csv for step 5.
Step 5 of 5
Write the cleaned copy and the notes its reader needs.
Recorded starter-catalogue results
Search
search: hand over a cleaned file its consumer can use
Download
download hand_
Placed where the harness reads it
Example layout
step-5-handover/ ├── customers.csv the task's file ├── addresses.csv from step 2 ├── merge-proposal.csv from step 4 └── .claude/skills/ └── hand-over-a-result-its-consumer-can-use/ └── SKILL.md the skill this step chose
Writes customers.cleaned.csv, the result of the task.
These searches reproduce the packaged starter catalogue identified above. They are recorded examples; a fresh search can return different results and versions.
Here each step takes the first result. In your own run, your agent reads the results and picks the one that fits. The folders show where each harness reads the skill, and how one step's output feeds the next.
When we measured items like these in real steps, they did not make a cheap model better at data cleanup, and one made it worse. Read the case study or see more examples.
A tabular competition, broken into six steps from the metric to the submission. Each step searches the library, downloads only the file it chose and puts it where the harness reads it.
Recorded starter-catalogue results, not a live search. The current library can return different versions. Always use the digest returned by a fresh search before downloading. About these recorded results.
The task
Predict the target of a tabular competition from train.csv, check that the score holds, and write submission.csv for test.csv.
Show the folders for
Step 1 of 6
Write down the competition's metric and the validation plan before any model runs.
Recorded starter-catalogue results
Search
search: freeze the success metric before measuring
Download
download freeze_
Placed where the harness reads it
Example layout
step-1-metric/ ├── competition.md the task's file ├── train.csv the task's file └── .claude/skills/ └── freeze-the-success-metric-before-measuring/ └── SKILL.md the skill this step chose
Writes validation-plan.json for the next steps.
Step 2 of 6
Check the training and test files for leakage, duplicates and label errors.
Recorded starter-catalogue results
Search
search: audit data splits for errors and leakage
Download
download audit_
Placed where the harness reads it
Example layout
step-2-splits/ ├── train.csv the task's file ├── test.csv the task's file ├── validation-plan.json from step 1 └── .claude/skills/ └── audit-data-splits-for-errors-and-leakage/ └── SKILL.md the skill this step chose
Writes folds.csv for the next steps.
Step 3 of 6
Fit a first model and read the gap between training and validation scores.
Recorded starter-catalogue results
Search
search: read the train validation gap
Download
download read_
Placed where the harness reads it
Example layout
step-3-baseline/ ├── train.csv the task's file ├── folds.csv from step 2 └── .claude/skills/ └── read-the-train-validation-gap/ └── SKILL.md the skill this step chose
Writes baseline-scores.json for the next steps.
Step 4 of 6
Group the validation errors by segment and look for repeated failures.
Recorded starter-catalogue results
Search
search: analyse errors by segment
Download
download analyse_
Placed where the harness reads it
Example layout
step-4-errors/ ├── baseline-scores.json from step 3 ├── folds.csv from step 2 └── .claude/skills/ └── analyse-errors-by-segment-and-cluster/ └── SKILL.md the skill this step chose
Writes error-report.md for the next steps.
Step 5 of 6
Rerun across seeds and folds, and keep only a result that holds.
Recorded starter-catalogue results
Search
search: check that a result is stable and generalizes
Download
download check_
Placed where the harness reads it
Example layout
step-5-stability/ ├── train.csv the task's file ├── folds.csv from step 2 ├── error-report.md from step 4 └── .claude/skills/ └── check-that-a-result-is-stable-and-generalizes/ └── SKILL.md the skill this step chose
Writes stability.json for the next steps.
Step 6 of 6
Predict the test file and write the submission with a short report.
Recorded starter-catalogue results
Search
search: hand over a result its consumer can use
Download
download hand_
Placed where the harness reads it
Example layout
step-6-submission/ ├── test.csv the task's file ├── stability.json from step 5 └── .claude/skills/ └── hand-over-a-result-its-consumer-can-use/ └── SKILL.md the skill this step chose
Writes submission.csv, the result of the task.
These searches reproduce the packaged starter catalogue identified above. They are recorded examples; a fresh search can return different results and versions.
Here each step takes the first result, and each folder shows where the harness reads the skill. This page shows how each step is set up, not a competition score.
See a data cleanup task step by step, or how a long task runs overnight.
Pro, $49 a month. Cancel any time.
We measured whether library items, placed in each step, help a small or cheap model clean data. On this test they did not, and one made results worse.
With the phone item, glm-5.3-flash:cloud scored 0.842 in all three repetitions, against 1.000, 0.974 and 1.000 without it. The item's digit-count rule is incomplete, and the model followed it: it lost every United Kingdom number written with a (0) trunk prefix and every number dialled with 011.
With the items, the cheap model spent 45 to 446 percent more prompt tokens per step in every family, and more requests and more seconds per step.
Without the items, the cheap cloud model already averaged 0.948 to 1.000 in each family and passed 12 of 12 steps. The larger glm-5.2:cloud was not separated from it in any family.
In all 37 of its steps, qwen2.5-coder:7b wrote its tool calls as plain text, so Pi ended each step with no output file. It scored 0 with and without material.
| Family | Without the item | With the item | Prompt tokens per step | Verdict |
|---|---|---|---|---|
| Phone numbers | 0.991, 3 of 3 passed | 0.842, 0 of 3 passed | 38,709 without, 87,038 with | Clearly worse |
| Email addresses | 0.952, 3 of 3 passed | 0.943, 3 of 3 passed | 71,927 without, 104,448 with | Not separated |
| Postal addresses | 0.948, 3 of 3 passed | 0.969, 3 of 3 passed | 17,446 without, 36,041 with | Not separated |
| Duplicate records | 1.000, 3 of 3 passed | 0.991, 3 of 3 passed | 22,135 without, 120,868 with | Not separated |
The cheap cloud model, glm-5.3-flash:cloud through Ollama Cloud, three repetitions each: record accuracy for phones, emails and addresses, and pair F1 for duplicates. Clearly worse means every repetition with the item scored below every repetition without it.
We stopped showing the phone item, and no demonstration on this site may use an item that a measurement found harmful.
It does not show that library material improves accuracy, lets a small model match a larger one, or saves tokens, time or money. It says nothing about data, models or harnesses that were not tested.
The 143 rows are synthetic, written by the study's author, who knew the items. The design was frozen before the first model call, each step ran three times in a fresh Pi process, and an independent scorer checked every output. The full report, every failure included, is in the public repository.
Can a cheap cloud model in an open-source harness use the library on its own? Gemma 4 31B in Pi searched the library, installed a skill, and a new Pi session loaded that skill and used it.
Pi has no built-in Model Context Protocol client, so Baltor connects through a Pi extension: one TypeScript file, served at /assets/pi/baltor.ts. It adds two tools, baltor_search and baltor_download.
Gemma 4 31B through Ollama Cloud. Pi calls the model with your own access; library downloads need no model key.
The file .pi/baltor.json names the environment variable that holds your Baltor key, never the key itself. The extension sends the key only to the service and removes it from every tool result.
The model searched the library with baltor_search. A search returns references, never file bodies.
baltor_download writes a file only when it is exactly the file that was published, so a changed or broken download never lands in your project.
The skill landed in .pi/skills/<name>/SKILL.md. The extension never replaces a folder it did not install.
A new Pi session loaded the skill. Pi lists a new skill from the next session on.
The model used the skill in that session.
Save the extension in your project, read it first, add the configuration and run the check. The check makes no model call and downloads nothing; its last line says Result: ready.
mkdir -p .pi/extensions curl -fsSL https://baltor.ai/assets/pi/baltor.ts -o .pi/extensions/baltor.ts pi -p --baltor-check
It was one check on one day. It shows the path from search to use working in Pi with a cheap cloud model; it does not measure whether the finished work was better than without the skill.
An account should belong to whoever controls its email address. This is how Baltor's sign-up makes sure of that, and what the live checks of September 24, 2026 showed.
On September 23, 2026 we probed our identity provider. Its own public sign-up kept the first password chosen for an address that was not yet confirmed. Whoever registered your address first could set its password, and you would activate their account by confirming it. Account creation stayed closed until the fix below was live.
accounts@mail.baltor.ai.Real accounts with disposable inboxes on the live service, at 13:03 UTC on September 24, 2026. No password, token or message body was recorded.
It came from accounts@mail.baltor.ai with the subject "Confirm your Baltor account", and its link opened the confirmation page.
Following the link confirmed the address, set the password and signed the owner in.
An address first registered through the provider's own sign-up was taken over by Baltor's sign-up. After that, the first password no longer worked, and the one chosen on Baltor's page did.
The identity provider's own public sign-up is still open at the provider. The service refuses every account it did not create, as the check showed.
Today the link in the message opens the service at its hosting address on fly.dev rather than at baltor.ai. Both addresses reach the same service.
Live readings from the service, taken each time you open this page.
Reading the service
This page asks the service now.
Search and downloads need every one of these to pass.
These cover the other parts of the service. A failure here does not stop it, but the part it names may not work.
This service only, as it answers right now: search, downloads, accounts and payment. Your model provider, your harness and your own machine are outside it.
When this page says the service works and your harness still fails, the troubleshooting page explains each answer the service gives.
Pricing
Agent Feeds is free through December 31, 2026. Harness Files adds the whole library for $49 a month, with the overnight tools included. Bring your own harness and model access.
01 · Research
$9 a month
Free through December 31, 2026 (Eastern)
Agent Feeds helps you choose tools and models with sources you can check. Start with source collections and the library updates feed.
No automatic charge. The standard price is $9 a month; a paid subscription requires your explicit opt-in.
02 · Build
$49 a month
Agent Feeds + Harness Files
While founding places last, a new account gets Pro free each month.
Subscribe from your account page once your account is ready.
03 · Keep working Public preview
$99 a month
Agent Feeds + Harness Files + Overnight / AFK Work
Overnight / AFK Work on a worker you control. The local tools remain included with Pro at no extra charge.
The run desk and local tools are free to explore today. Max subscriptions are not sold during this public preview. Your worker uses your own model access and hosting.
No. A Baltor key gives access to the library. The model calls happen in your own tools, with your own provider account or your own models.
They stay where you saved them. Revoking a key or cancelling stops new requests; it does not recall files that were already delivered.
Yes. Agent Feeds is free through December 31, 2026 (Eastern), with no automatic charge. A paid feed subscription will need your opt-in. The Public Good collection is free with an account. The Baltor Harness is free and open source. Agent Feeds + Harness Files adds the full library for $49 a month. While founding places last, a new account gets Pro free each month. Your account holds a founding place, so Agent Feeds + Harness Files is free for it each month. Your account includes Pro free each month.
Not today. One account holds one person's keys. Each person on a team needs their own Agent Feeds + Harness Files account.
Create your account, subscribe to Pro for $49 a month and connect your harness.
Step 1 of 5
Enter your email. We send you a link to confirm it.
Step 4 of 5
Public Good downloads are free with an enabled account and fair-use limits. Pro is optional for the wider library.
Already have an account? Sign in
Step 3 of 5
Your email address is confirmed when you set it. Then you are signed in.
It has expired, it was used already, or a newer message replaced it. Nothing was changed.
Ask for a new linkYour account is open.
Sign in with your email address and your new password.
Sign inYour workspace
Sign in to your account.
Never enter your model key here. It stays in your own tool.
This form checks real service access; it does not create an account or a subscription.
Accounts
It takes a minute: your email address, a link to confirm it and a password you choose.
Get started walks you through creating your account, subscribing and connecting your harness.
Enter your email address. We send you a link, and you choose your password on the page it opens. Creating an account does not start a subscription or give access to your local files.
Public Good files are free with an enabled account, within request and byte limits. No paid plan is required for those files.
Get started
Leave your email address, and we send your link to finish setting up.
Accounts cannot be created on this service right now, so no form is shown and nothing you type here is recorded. To ask for an account, write to us from the support page.
Your first subscription may come with a discount code, and checkout takes that code.
We keep your address and note only to send your link, as the privacy notice says, and we can erase both on request. We will not add you to a mailing list.
Already have an account? Sign in
Workspace
Check a result's source and licence, then download the exact version. Your own tools run in your environment.
Search
Search returns references. Fetching a body requires a separate access check and may record usage.
Connect, describe the material you need, then inspect the returned source and permissions before downloading.
Staff read every request. Write only what a step needs; the text is kept with your account.
The library
Search answers a question. The table shows everything that is there: skills, instruction files, rules, subagents, commands, hooks, plugin manifests, protocol server settings, schemas and code. Type to narrow it, sort a column, and open a row to read its details before you fetch the file.
Tool filters include matching tags and files whose compatibility is not recorded. Every file your account may use is listed with the effects it declares. Before you check or fetch a file that declares more than reading files, such as running a command, the page names those effects and asks you to confirm. Nothing runs on this page.
Sign in to browse the library published for your account.
Your account
Your plan, the Baltor keys your tools use and what they downloaded, as the service records them.
Sign in to see your service identity, usage and available subscription settings.
Your account includes Pro.
One Baltor key for each tool or device. It can search and download what your plan allows. It cannot manage your account, your billing or other keys.
Sign in with your verified account to manage client tokens.
Choose what this connection may do and when it expires. It cannot manage accounts, billing or other credentials.
Shown once. Closing or reloading the page clears this copy. The service keeps only a digest of it, so nobody can show it again.
Next, add it to your harness: open the settings for your tool.
Signing out of this website does not revoke these keys. Revoke them here when a device no longer needs access.
Open the connection settingsGive each bot a limited assignment. It can pick up work, checkpoint and submit candidates through the shared work API or MCP. It cannot publish, run uploaded files, manage accounts or spend on models through this service.
Sign in with your email to manage worker assignments.
First create a key above with only “Search material and send library feedback” selected. Keep that key in the bot's secret settings, not its prompt. Select it here to grant bounded work access.
Disabling an assignment blocks new work requests. It cannot recall data already read or stop actions in the bot's own environment.
Tasks stay in your account. Creating a proposal does not make it claimable. Inspect its exact version, then admit it. External tools and model usage still need authority in the bot's own environment.
UTF-8 candidate files are limited to 16 files, 64 KiB each and 256 KiB combined, subject to the smaller encoded request limit shown above. Binaries and media uploads are not supported here. Accepted work is still not a published library component. No automatic retry or expired-lease requeue is performed.
Save a feed, preview what it would send, then choose which Baltor key reads it. Each key reads only the feeds you assign to it. Pause or remove any of them here.
Sign in with your email to save feeds and choose which key reads each one.
Start from a topic or a collection of sources. Your agents receive the sources, links and checked dates, in the format you choose.
Feed text comes from outside sources. Your agents receive it as data to read, never as instructions to follow.
Each item your tools downloaded, how many times and when last. Service usage is separate from the model calls your own tools make.
Sign in to see the downloads recorded for your account.
New Pro subscriptions are $49 a month. Existing subscribers keep their subscribed price, which the customer portal shows. Card details stay with Stripe. Choose a plan here, or open the customer portal to change or cancel it at any time. Only configured plans are shown, and a completed checkout does not itself confirm that access is active.
Staff
Accounts, customer feedback and test tokens, read from the service records. The service checks every permission again.
Sign in with your staff account or an authorized administrator service token.
Share research, test results and useful files. Read a saved report to reply.
Task brief · Feedback brief · Privacy notice
Tokens share the selected tenant's material and usage. A token is not a separate private account.
Shown once. It cannot be recovered from the service. Store it in your secret manager.
Revocation takes effect on the next service request. It does not recall material already downloaded. Bootstrap administrator keys are managed separately.
Staff
Recorded release, catalogue and edge checks, each with its source revision, environment, attempt, time and record, beside what the running service measures now. The service checks the staff role on every request.
Sign in with a staff account to read this view. The service checks the role on every request.
Sign in to open SystemsFailed, pending, skipped, unknown and passed checks are listed separately, newest record first, a page at a time. A row whose environment is not stated, or that ran locally, is not production verification.
Earlier releases, publications and attempts stay listed. Open an entry to read every row it holds.
How it works
Start with what you want to achieve. Break the work into steps, give each step relevant expertise and tools, then check the result before moving on.
The plan can change as you learn more. A step can ask for missing information, try another approach or send unfinished work back for improvement, within the limits you set.
See what each step needs
Choose a step to see its information, tools and expected result. The person inspecting the data needs different material from the person checking the finished import.
This is an illustration, not a recorded customer result. Selecting a step does not access your files or call a model.
Useful information, when it is needed
A useful briefing is not always the shortest one. Start with relevant information and add more when the work reveals a gap.
It can search for additional material or ask a question. Your permissions and remaining budget still apply. New instructions cannot grant access to private files or allow spending by themselves.
If an action times out, first check whether it happened. Trying again must not send a second email or repeat another change by mistake.
Use what suits the work
The goal and quality checks stay the same. The way a step does the work can vary.
Give your chosen development tool a clear brief, selected files and the permissions needed to build or repair the missing pieces.
A decision may need one model request rather than a complete development session. Use a compatible model service you have already configured.
Use existing tested code when it meets the requirements, saving model calls for the steps that need new code.
Models run on your machine or through a provider you choose, using the model access you already control.
Make the effort count
These are the questions Baltor is designed to address at each step of the work.
Match information to the task. Bring in more when it helps, without making every step carry the whole history.
Relevant information, less repetitionCompare suitable models and reusable code for each step, against your quality and spending requirements.
An approach that fits the workBring relevant methods, examples and checks into the work, with sources and applicability kept visible.
Specific guidance, not generic guessesFind an eligible implementation first. Generate new code for the gaps instead of reproducing a known solution as more output tokens.
Reuse what is already qualifiedKeep useful results and corrections. Compare proposed improvements on real tasks before making them the default.
Measured improvement, not blind repetitionBuilding blocks for better solutions
Useful intelligence includes how to approach the work, code that can do it, what happened before, and what you want done differently.
Methods, questions, constraints, examples and output contracts that guide the work.
A field-definition guide and a checklist for missing values.
Reusable functions, tools, files and workflows with declared inputs, outputs and effects.
A qualified normalizer with its dependency and verification records.
Saved outcomes, failures, repairs, measurements and solution information that can inform a new task.
A previous import's failed assumptions and the checks that detected them.
Scoped corrections, priorities and instructions supplied by a person.
“Keep uncertain matches for review. Never overwrite the original file.”
The four layers describe the broader product. The hosted library serves reusable files and reference material. Coverage and execution evidence vary; a populated library is not proof of useful results on every task.
Your tools, with help from Baltor
Anyone can create an account on Get started. This service is not taking new accounts right now. With an account, you sign in with your email address and password, create a Baltor key for each of your tools on your account page, search the library, download the files your plan includes and see your usage.
The complete example above is still being tested from start to finish. The playable browser demo supports scene changes and editable export. We have not established token savings or overnight task completion.
Free Public Good components, beginning with worker protection, are available on Public Good. See the next product steps in the public pitch deck.
Get set up
Three steps: create your account, connect your harness, make your first download. Your harness is the program that runs your coding agent, such as Claude Code, Codex or Cursor. Your tools and model access stay yours.
Use one, or all three. None of them needs your model key.
MCP server
One server entry in Claude Code, Codex, Cursor, OpenCode, VS Code, Pi or any app that speaks MCP (Model Context Protocol). Your agent searches the library and downloads exact files.
Checking the connection address
OAuth sign-in in apps that support it, or a Baltor key from your environment in any app.
Choose your harnessREST API
For CI jobs and your own harness. Send a Baltor key in the Authorization header, search, then download the exact version you chose.
POST /api/v1/retrieval
Loading the example request
Feed links
No account and no key. Put a source collection's link in your agent's instructions, as JSON Feed or Markdown. Library updates also come as RSS.
/feeds/collections/coding-benchmarks.md
Step 1
Leave your email address on Get started, and we send your link to finish setting up.
Get startedConfirm your email address and choose a password. Creating an account does not start a subscription.
Get startedGet started shows the account options this service offers. For help with an existing account, use the support page.
Already have an account? Sign in
Step 2
Baltor accepts OAuth sign-in from Claude Code, Claude on the web and Claude Desktop, Cursor, VS Code, Codex and OpenCode. In the app, add and choose its sign-in option. Sign in to Baltor, review the app and permissions, then allow the connection. Your password and model keys are not handed to the app.
codex mcp add baltor --url Checking the connection address. Codex starts the Baltor sign-in itself; run codex mcp login baltor only if that sign-in did not finish.opencode mcp auth baltor.In Claude Code, run these two commands in your project folder. The second one opens the Baltor sign-in page in your browser:
Loading the connection commands
The app needs authorization-code support with PKCE and resource-bound tokens. Connection details and limits. The key settings below work for every harness listed, including apps that cannot sign in with OAuth.
Download the container with OpenCode and the Baltor harness. Bring your own model access.
Pick yours below. Create a Baltor key, keep it in your environment, and copy the connection settings. The configuration contains the variable name, never your key.
For apps such as Open WebUI, use the OpenAPI connection to search the same library, inspect an exact version and download its files.
https://baltor.ai/api/v1/connect as the server address.request_id.Inspect the OpenAPI document. Search and manifests read metadata. The first download of an item version in a calendar month counts once. Files with additional effects still require the step's declared permissions.
Loading the connection settings
Merge this entry into your existing configuration. Do not replace your other settings.
Loading the secret-free configuration.
No key included. Set BALTOR_SERVICE_TOKEN to your Baltor key in your environment. These settings contain the variable's name, never your key. Check your setup with Choose a client above.
Loading documented compatibility information.
These apps reach the same address with a Baltor key. Each entry holds a reference to your key, never the key itself. For Cursor and Gemini CLI, start the app from a terminal where BALTOR_SERVICE_TOKEN is set.
.cursor/mcp.jsonAdd this entry to .cursor/mcp.json in your project, or to ~/.cursor/mcp.json for every project.
Loading the connection settings
Check it in Cursor Settings under MCP, or run agent mcp list with the Cursor CLI.
.vscode/mcp.jsonAdd this to .vscode/mcp.json in your workspace. VS Code asks for your Baltor key once and keeps it in its own secret storage.
Loading the connection settings
Check it with MCP: List Servers in the Command Palette.
.gemini/settings.jsonAdd this entry to .gemini/settings.json in your project, or to ~/.gemini/settings.json.
Loading the connection settings
Check it with gemini mcp list.
Any other app that speaks MCP over Streamable HTTP connects the same way: the address above, your Baltor key in the Authorization header from your environment, and Baltor-Step-Effects to say what its steps may do.
This prompt hides the value while you enter it. The value is not part of the command history. The client process inherits the environment variable.
read -rsp "Baltor key: " BALTOR_SERVICE_TOKEN export BALTOR_SERVICE_TOKEN
After closing the client, clear the terminal variable with unset BALTOR_SERVICE_TOKEN. Do not run environment dumps or record the terminal while handling secrets.
Step 3
Ask your connected agent to find files for a real task. For example:
Search Baltor for a CSV validation tool. Show its source, licence and requirements before downloading it.
Your agent sees results first. It downloads the files you chose with their required dependencies; the download appears in your usage.
Connecting needs nothing installed beside your harness. The Baltor Harness runs on your machine and is free and open source under the MIT licence; the documentation has its installation steps.
A fantasy arena with original characters, sword combat and editable animation. Open Scene tools to export a GLB for Blender.
An original example, not a measured comparison with a model working without Baltor. Source and setup
Watch a task run step by step, read what happened in four real cases, and try a first search yourself.
Demonstration, 5 steps
Watch a data cleanup task run step by step. Each step searches the library, downloads only the file it chose and puts it where the harness reads it.
Open the demonstrationDemonstration, 6 steps
A tabular Kaggle competition in six steps, from fixing the metric to writing the submission file.
Follow the competitionReal results, including what did not work.
In your own workspace: search, look closer, then download only what you chose. No model call is needed.
Sign in, then search for review inputs. You get short results, not whole files.
This prepares the query in your workspace. You choose when to search.
Open Source, integrity and access on a result to see where it comes from, its licence and its exact version.
Select Download this version to download that exact version. The download counts in your usage.
Open your account to see the download in your usage. If a download times out, try the same selection again.
To see a whole task, from the first search to the result, open the demonstration.
Pro, $49 a month. Cancel any time.
Connecting to Baltor grants access to the files your plan includes. It does not give the service permission to run commands on your computer.
Your Baltor key reaches the library. Your model key stays in your harness, and Baltor never asks for it.
A downloaded file grants no permission to run anything. Your harness decides, with the permissions you set.
Search returns references. A file arrives only through a download that names the SHA-256 of its exact bytes.
An item that declares effects on your computer, such as reading or writing files, says so before you download it, and your harness sets what each step may do.
A Baltor key identifies one connection with limited permissions. The service stores key digests, checks scope and expiry, and can revoke access. Your model provider keys belong with your local client or approved credential broker. Do not enter them into the website.
An email sign-in lasts as long as your browser tab: the page keeps it in the tab's session storage, so reloading keeps you signed in, and closing the tab or signing out ends it. A Baltor key entered on the sign-in page stays in page memory only, and reloading clears it. Avoid shared devices, untrusted extensions and screenshots of credentials.
Search returns permitted metadata. Fetching a body rechecks access and its exact identity. A digest proves that bytes match the selected reference; it does not prove that the material is safe, correct or useful.
Inspect unfamiliar code and plugins. Use a confined workspace and a sandbox appropriate to the work, with explicit file, command and network permissions. A shared container does not isolate one process from every other process inside it.
Every admitted item has an approval record tied to its exact bytes. No reviewer judges an item it wrote, and one written rejection keeps an item out, with the reason recorded.
Full per-item review needs approval from at least two reviewers of at least two model families, none of them from the family of the model that wrote the item. An initial screen uses one calibrated independent reviewer against written criteria: the practice it describes is sound, what it says about its cited source is true, its licence is settled, and it declares every effect its steps have on your computer, such as reading or writing files.
Generated batches use per-item automated checks and independent sampled review. Since October 5, 2026 a generated item is published once it passes the automated checks, and the sampled review continues after publication and withdraws what it rejects. A calibrated reviewer from a different model family than the generator's author reviews a random sample, with deliberately defective examples included to check the reviewer. Not every item in an accepted batch receives a model review. Its evidence records which checks ran and whether its independent review is complete or ongoing. Approval does not imply installation or execution, or an improvement in your task results.
Before any reviewer reads an item, automated checks refuse it for a licence that is not accepted or that disagrees with the licence the item declares, a missing part, unsafe instructions such as a download piped into a shell or a read of a credential file, declared effects that do not match its steps, a value shaped like a credential, or a copy of another item. Material from outside sources also needs a record of where its exact bytes came from.
The approval record names each reviewer, with its decision and its reasons. The library's first 43 items were approved on September 21, 2026, before the rule on model families: three reviewers who wrote none of them, one for correctness and usefulness, one for provenance, licence and safety, and one adversarial, each approved every one of those items, and 6 of the 49 items they judged were rejected. One of those reviewers came from the same model family as the model that wrote the items.
Approval gives an item no permission to run anything. The digest proves that the bytes you download are the approved bytes; it does not prove that they suit your task.
Search text, requested references and service authentication reach the server when you use the workspace. The service records access and usage metadata. The website does not automatically upload your project files, model keys or complete execution traces.
You choose what to send through available submission and upload features. Send content you have permission to provide, check the audience and retention information, and leave out credentials and unrelated personal data. Read the privacy notice for storage, access and deletion details. Embeddings and derived features are not automatically anonymous.
The service runs on one Fly machine in one region with an attached persistent volume. A local backup-and-restore exercise has passed. This is not a multi-region service, an availability guarantee, an independent security audit or a compliance certification.
Anyone can create an account on Get started: you give your email address, open the link we send and choose a password. This service is not taking new accounts right now. You sign in with your email address and password. Baltor keys for your tools are created and revoked on your account page, and a Baltor key cannot manage your account, your billing or other keys.
Read the privacy notice for what the service stores, and set the permissions of each task in your own harness. Keep provider credentials out of prompts and retrieved files. Do not repeat an external action after an uncertain outcome until its state has been reconciled.
Pro, $49 a month. Cancel any time.
Privacy
Kind: privacy notice. Updated October 1, 2026 with the owner's approval of voluntary submissions, uploads and data exchange for service operation and improvement. Available features and account permissions determine which operations a person or connected tool can use.
Operator: Baltor.AI, 1428 Bryn Mawr St, Saxton, PA 16678, United States.
Baltor supplies reusable material and tools and supports research, feedback, testing and collaboration. It processes the inputs needed for the operations you request. Browsing and downloading material do not give Baltor automatic access to your computer, local files, prompts or model keys.
You or a tool acting with your permission may send messages, feedback, research, source links, reports, files and other inputs through available forms, uploads, APIs, command-line tools or connections. Baltor may receive, read, store, organize, analyze and return that content to provide requested features, respond to questions, investigate problems, evaluate quality and improve its services and reusable materials. Send only content you have permission to provide, and leave passwords, access keys and unrelated personal information out of submissions.
Where an integration is available, it may exchange selected inputs and results with the service or tool you connect, within the permissions you grant. Private submissions do not become public library material just because they were uploaded. Publication and reuse require the relevant permissions and review. This notice does not grant access to an account, authorize executing uploaded code or make every described interface available to every account.
| Data | Why | Where |
|---|---|---|
| Your email address and password | To sign you in. The password is kept only by the identity provider, in hashed form. Baltor never stores or sees it after sign-up. | Supabase, United States |
| An account record that links your sign-in to your account | To know which material you may read | The service database on Fly, United States (iad) |
| A digest of each access key | To check a key without being able to show it again. The key itself is shown once and never stored. | The service database |
| One usage record for each downloaded item: the item name, its digest and the time | To show usage and reconcile access | The service database |
| Feedback, ratings, requests and other content you explicitly submit, together with the account, dates and version or request references needed to understand it | To respond, reproduce problems, identify missing material and improve the service | The service database on Fly, United States (iad) |
| Submitted reports and files, source links, replies, checksums and immutable revision records | To support collaboration, review, reliable retrieval and improvement work | The service database and private file storage on Fly, United States (iad) |
| Connection and consent records, client metadata, granted scopes, expiry information and token digests | To operate authorized connections, check their permissions and investigate access problems | The service database |
| A digest of a browser session that you signed out of, until it would have expired | To refuse that session afterwards | The service database |
| A waiting list entry, if you ask to join the list: your email address, an optional note of at most 280 characters, the state of the entry and its dates | To invite people in small groups | The service database |
| Diagnostic records of refused requests: a random reference, address path, reason code, account when known, time and service version; request content only when detailed capture is explicitly enabled | To find and fix problems. The default retains the latest 500 failures with metadata only; configured capture and retention limits apply. | The service database |
| Your payment details when you use an available billing feature | Stripe collects and keeps payment details. Baltor receives the customer identifier and subscription information needed to manage access. | Stripe |
The service keeps a count of refused sign-in attempts for each network address in memory for a short time, to slow down guessing. It is not written to disk. To stop one machine from flooding the waiting list, the service also keeps the times of recent waiting list requests under a keyed one-way digest of the sending network address, never the address itself, and removes them once the one-hour counting window has passed.
The default diagnostic profile does not store request bodies or network addresses. Explicitly enabled detailed diagnostics may retain supplied request content for troubleshooting within the configured limits. That content can include project or personal information present in a request. Diagnostic records are restricted, not a public feed; do not rely on automatic removal of every sensitive detail before submitting content.
Access to private submissions follows the account and feature permissions. Authorized service operators may review submissions for support, security, quality evaluation and improvement. Content displayed in a private work log is not an anonymous public feedback feed. Removing a name alone may not make a report anonymous; public summaries need review for identifying details.
Hosting, identity and payment providers process the information needed to run their respective parts of the service. Optional connected tools or model providers may process content when that workflow is enabled and authorized; their own terms and data practices also apply. Uploading a file by itself does not trigger a model call or send it to an unrelated external service. Baltor does not sell personal information. A connection or this notice does not authorize unrelated advertising disclosure or unrestricted access to other accounts.
The hosting provider keeps daily snapshots of the service database for five days. Feedback, reports, uploaded files and their revision history currently have no automatic expiry. They remain available for the purposes above until an authorized removal; deleting a browser session or ending a connection does not erase the records it created. Token expiry limits access, not the lifetime of the associated registration or consent history.
To request deletion of your account or submitted content, write to the contact address below and identify the relevant account or report. Requests are handled by an operator; self-service deletion is not available. Removal must account for stored copies, attachments and revision history, not just hide a listing. Records needed for security, disputes or applicable legal obligations may require separate retention. To leave the waiting list, ask in the same way: the entry keeps only a one-way digest of the address and the history of decisions, so the address itself is gone. Database snapshots expire within the provider's five-day window.
Baltor.AI, 1428 Bryn Mawr St, Saxton, PA 16678, United States.
Questions that contain no personal data can also go to the public issue tracker of the repository. Do not post personal data in a public issue.
This notice changes when the service changes what it stores. The history of this file is the record of those changes. A revised notice does not by itself authorize a materially different use of information collected under earlier promises. Additional notice or consent is needed where applicable.
Terms
Operator: Baltor.AI, 1428 Bryn Mawr St, Saxton, PA 16678, United States.
Last changed:
The privacy notice says what the service stores about you.
Documentation
Connect your own tools, find files for each task and check your downloads.
One page for each harness: what you need, the configuration to paste, the first search, the first download and the check that it worked. Claude Code, Codex, OpenCode, Pi or the Baltor Harness.
What components, packages and files are, how to handle updates and withdrawals, and answers to common questions.
Loading the documentation list.
The local engine is free and open source under the MIT licence. Its quickstart uses Python 3.10 or newer and Docker for code execution. Connecting your existing harness to the hosted library does not require this installation.
python3 -m venv .venv source .venv/bin/activate python -m pip install "https://github.com/alisonjieli-png/loop-engine/archive/refs/heads/main.zip" loop-engine doctor
The engine calls the model you choose, such as a local Ollama, Ollama Cloud or a model server on your own network, with your own key. The repository README has the full steps, including Windows and macOS.
The public website calls a unit of work a step. The implementation uses the canonical Loop runtime described here and in the repository.
A discrete cognitive or act step Loop node is an independently governed instance of the Loop runtime responsible for one clearly defined cognitive step or action. A cognitive step might interpret information, identify a missing requirement, compare alternatives, or evaluate a result. An action might inspect a directory, build software, execute a test, create an artifact, or send an authorized email.
Each discrete cognitive or act step Loop node receives the context, instructions, skills, plugins, tools, and working files relevant to its assignment. Essential information can be supplied directly, while additional information can remain in centralized storage behind authorized, versioned references. It does not automatically need the entire task history or every available tool.
A separately initialized harness process, such as OpenCode, Pi, Codex, or a custom implementation, can perform the assignment. When explicitly permitted, another harness can attempt the same assignment after a failure. The assignment's contracts, permissions, history, and remaining authority persist across those attempts.
Discrete describes the scope of the assignment, not a restriction to one attempt, one model call, or one output. A discrete cognitive or act step Loop node can examine whether an observation matches its expectations, identify a problem, repair or change its approach, and repeat until its declared completion conditions are satisfied.
Alternatively, a discrete cognitive or act step Loop node can publish an initial candidate output and continue working while its continuation conditions and authority permit. It can produce additional alternatives over time, including alternatives that are better, worse, or useful under different circumstances. Consumers must identify exactly which output they used. Publishing an output does not necessarily mean that the producing assignment has finished.
For externally consequential actions, continued operation does not authorize repeated effects. For example, generating alternative email drafts can continue, but sending an email requires its own authorization and protection against duplicate delivery.
Harness Intelligence is a provisioning view over the four persistent layers. It is not a fifth persistent layer.
Runtime Memory is the temporary note board for one run. Saving a note does not promote it into reusable intelligence.
Through your signed-in connection
With a direct connection, your model key stays in your environment or its approved credential manager. A remote model may receive the information you allow your tools to send. With the model proxy, Baltor uses the provider key for that request and does not store it.
Every executable graph vertex is a Loop. A harness is an implementation adapter used by a Loop. Files, services, contracts and stores are not additional executable vertices.
Operational runtime type
└── Loop
├── Operational relationship
│ ├── Starting
│ ├── Spawned by
│ ├── Queried by
│ ├── Retrieved by
│ └── Connected from
├── Role
│ ├── Practitioner
│ ├── Intelligence
│ └── Solution
├── Versioned role profile
├── Purpose and domain categories
├── Run mode
│ ├── deterministic
│ ├── hybrid
│ └── non-deterministic, with model-led semantic work
├── Step profile
├── Typed input and output contract
├── Loop condition
├── Exit condition
├── Graph relationships
├── Budget, permissions, and effect policy
├── Model settings when the selected mode permits a model
└── Run History recordsLoop role profiles
├── Practitioner
│ ├── reference nine-step
│ ├── compact five-step
│ ├── research
│ ├── solver
│ ├── verifier
│ ├── code execution
│ └── self-improvement task
├── Intelligence
│ ├── cross-layer search and materialize
│ ├── Context Intelligence: serve, search, and frame
│ ├── Code Intelligence: resolve, invoke, and load
│ ├── Runtime History and Solution Intelligence: search, replay, and compare
│ └── User Feedback Intelligence: serve, scope, and interpret
└── Solution
├── atomic component
├── pipeline
├── router and fallback
├── ensemble
└── validator