The perfect harness working directory for each task or subtask.

Your agent searches a vetted library, and Baltor places the instructions, skills, tools and settings it chose where your harness reads them. No manual search, no manual setup, nothing copied by hand.

Get started Get set up

Baltor Pro $29 a month for the whole library

Claude Code
Codex
OpenCode
Pi
Baltor Harness

The working directory for one step

Example layout

step-3-find-duplicates/
├── AGENTS.md                     only this step's context
├── CLAUDE.md                     imports @AGENTS.md
├── .claude/
│   ├── settings.json             this step's permissions
│   └── skills/                   5 vetted skills it chose
│       ├── find-duplicate-records-with-blocking-keys/
│       │   └── SKILL.md
│       ├── score-duplicate-pairs-by-weakest-signal/
│       ├── canonicalize-company-legal-suffixes/
│       ├── normalize-and-recover-email-addresses/
│       └── propose-a-dedupe-without-deleting-rows/
├── .mcp.json                     its tools, nothing more
├── lib/
│   └── blocking_keys.py          reused code, not rewritten
└── .baltor/step.lock.json        versions and digests

The local engine is built to assemble a directory like this on demand for every step of a task.

See it work

Start to finish, for three kinds of work.

A simple task
Clean a customer file
Five small steps. Each one searches the library, downloads only the file it chose and shows where the harness reads it.
Watch the five steps
A long task, overnight
Run a long job while you sleep
The local engine is built to run a big job in small steps, each started with only its own files and checked before the next builds on it.
How a night of work runs
A Kaggle solution
From the metric to the submission
Six steps of a tabular competition, from fixing the metric to writing the submission.
Follow the competition

The library

Anything a harness can pick up, in one place.

Skills with their scripts, instruction files, tools and code, subagents and commands, hooks and protocol server settings, searched from inside your harness. New vetted additions keep arriving.

Skills
SKILL.md · scripts/ · assets/
Instruction files
AGENTS.md · CLAUDE.md
Tools and code
tested scripts and packages
Subagents and commands
agent and command definitions
Hooks
checks your harness runs itself
Protocol servers
the tools a step may call

What it makes possible

Three things teams do with Baltor.

One account holds one person's keys, so each person on a team holds their own account.

Solve complex problems overnight
Hand a big job to a small or local model before you log off. The Baltor Harness breaks it into small steps and gives each one only the files it needs, so the work keeps moving while you sleep.
How overnight work runs
More efficient operation
Built so each step carries only the context it needs, reuses proven code instead of writing it again, and runs on the smallest model that does the job well.
Where the savings come from
Learning and optimization, built in
Every run leaves a record of what worked, what failed and which setup was fastest. Baltor is designed to use those records to set up the next task better.
How it keeps improving

Built for teams that check the details.

Your keys stay yours
Baltor never asks for a model key. Your harness calls your models through the connection you control. One account holds one person's keys.
Exact versions
Every download names its exact version, so your agent works from the file that was reviewed, not a later edit.
Nothing runs on your computer
Connecting gives Baltor no permission to run commands on your computer. Your harness decides what runs, with the permissions you set.

Pricing

One plan. The whole library.

Baltor Pro is $29 a month. You bring your own model access, so model costs stay with your own provider, and there is no overage billing.

Baltor Pro

One plan

$29a month

The whole library, searchable from your harness
One downloaded item is the measured unit, with no overage billing
A key for each client you connect, issued to your account
Usage you can see, item by item

Get started

Cancel any time from your account page.

See what is included

Questions, answered.

Anything else, write to the operator at the contact address in the privacy notice.

Which harnesses does it work with?

Claude Code, Codex, OpenCode, Pi and the Baltor Harness. Get set up shows the exact steps for each. Another client can connect when it speaks the Model Context Protocol version this service accepts and sends the key in a header.

Do I need to run a model myself?

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. Baltor never asks for a model key.

What counts as a download?

One item at the version you selected counts once. A search, a second look at the same reference and a refused download are not downloads.

What does Baltor keep about my work?

For each item your agent downloads, one usage record: the item's name, its digest and the time. Your tasks, files, prompts and model keys stay on your computer. The privacy notice lists everything the service stores and for how long.

Give your agent the right setup for every task.

Search the whole library from the harness you already use.

Get started Get set up

Savings depend on the task, the model and the setup.

Demonstration

One task, five small steps.

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.

Get started Get set up

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.

Profile the columns
Split the addresses
Find duplicates
Propose merges
Hand over the copy

Step 1 of 5

Profile the columns

Count what each column holds before anything changes.

Real results from the library

Search

search: profile a text column before cleaning
1. profile_text_column_before_cleaningskill · MIT · 2.9 KB · sha256 3274cbf5…Chosen
2. layer_exception_catalogs_with_precedenceskill · MIT · 2.8 KB · sha256 0a8cc8cd…
Download
download profile_text_column_before_cleaning

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

Split the addresses

Break each address line into number, street, city, region and postal code.

Real results from the library

Search

search: split address lines in a customer file
1. split_address_lines_into_componentsskill · MIT · 3.0 KB · sha256 53dc74e3…Chosen
2. find_duplicate_records_with_blocking_keysskill · MIT · 2.7 KB · sha256 34da79f4…
Download
download split_address_lines_into_components

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 duplicates

Find the customers entered more than once.

Real results from the library

Search

search: find duplicate customer records
1. find_duplicate_records_with_blocking_keysskill · MIT · 2.7 KB · sha256 34da79f4…Chosen
2. score_duplicate_pairs_by_weakest_signalskill · MIT · 2.6 KB · sha256 9e713a45…
Download
download find_duplicate_records_with_blocking_keys

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 merges

Propose one survivor for each group of duplicates, without deleting a row.

Real results from the library

Search

search: propose a dedupe without deleting rows
1. propose_a_dedupe_without_deleting_rowsskill · MIT · 2.6 KB · sha256 c287344c…Chosen
2. find_duplicate_records_with_blocking_keysskill · MIT · 2.7 KB · sha256 34da79f4…
Download
download propose_a_dedupe_without_deleting_rows

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

Hand over the copy

Write the cleaned copy and the notes its reader needs.

Real results from the library

Search

search: hand over a cleaned file its consumer can use
1. hand_over_a_result_its_consumer_can_useskill · MIT · 2.7 KB · sha256 ddef271b…Chosen
2. orient_on_a_task_and_write_its_contractsskill · MIT · 2.9 KB · sha256 13c13dbc…
Download
download hand_over_a_result_its_consumer_can_use

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.

Real searches

Every search on this page, and the results it returned in that order, come from the library as it is today.

Your agent chooses

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.

What a measured run found

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.

Demonstration

A Kaggle competition, start to finish.

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.

Get started See a data cleanup task

The task

Predict the target of a tabular competition from train.csv, check that the score holds, and write submission.csv for test.csv.

Fix the metric
Audit the splits
Train a baseline
Find where it fails
Check the score holds
Write the submission

Step 1 of 6

Fix the metric

Write down the competition's metric and the validation plan before any model runs.

Real results from the library

Search

search: freeze the success metric before measuring
1. freeze_the_success_metric_before_measuringskill · MIT · 2.8 KB · sha256 8bb77dda…Chosen
2. orient_on_a_task_and_write_its_contractsskill · MIT · 2.9 KB · sha256 13c13dbc…
Download
download freeze_the_success_metric_before_measuring

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

Audit the splits

Check the training and test files for leakage, duplicates and label errors.

Real results from the library

Search

search: audit data splits for errors and leakage
1. audit_data_splits_for_errors_and_leakageskill · MIT · 2.7 KB · sha256 9c2133a4…Chosen
2. analyse_errors_by_segment_and_clusterskill · MIT · 2.6 KB · sha256 d08d4555…
Download
download audit_data_splits_for_errors_and_leakage

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

Train a baseline

Fit a first model and read the gap between training and validation scores.

Real results from the library

Search

search: read the train validation gap
1. read_the_train_validation_gapskill · MIT · 2.8 KB · sha256 5decc107…Chosen
2. audit_data_splits_for_errors_and_leakageskill · MIT · 2.7 KB · sha256 9c2133a4…
Download
download read_the_train_validation_gap

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

Find where it fails

Group the validation errors by segment and look for repeated failures.

Real results from the library

Search

search: analyse errors by segment
1. analyse_errors_by_segment_and_clusterskill · MIT · 2.6 KB · sha256 d08d4555…Chosen
2. audit_data_splits_for_errors_and_leakageskill · MIT · 2.7 KB · sha256 9c2133a4…
Download
download analyse_errors_by_segment_and_cluster

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

Check the score holds

Rerun across seeds and folds, and keep only a result that holds.

Real results from the library

Search

search: check that a result is stable and generalizes
1. check_that_a_result_is_stable_and_generalizesskill · MIT · 2.7 KB · sha256 6ba31af0…Chosen
2. check_for_existing_work_before_buildingskill · MIT · 3.0 KB · sha256 b710af36…
Download
download check_that_a_result_is_stable_and_generalizes

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

Write the submission

Predict the test file and write the submission with a short report.

Real results from the library

Search

search: hand over a result its consumer can use
1. hand_over_a_result_its_consumer_can_useskill · MIT · 2.7 KB · sha256 ddef271b…Chosen
2. check_that_a_result_is_stable_and_generalizesskill · MIT · 2.7 KB · sha256 6ba31af0…
Download
download hand_over_a_result_its_consumer_can_use

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.

Real searches

Every search on this page, and the results it returned in that order, come from the library as it is today.

Your agent chooses

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.

More demonstrations

See a data cleanup task step by step, or how a long task runs overnight.

Case study

Data cleanup with and without Baltor.

We measured whether approved 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.

Measured September 22 to 23, 2026 · 143 synthetic rows · 72 counted steps · Pi 0.73.1

Get started See a task step by step

0 of 4families where the item was clearly better
1 of 4clearly worse: phone numbers
45 to 446%more prompt tokens per step with the item
12 of 12steps the cheap model passed without any item

What we found

Worse on phone numbers
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.
More tokens, more time
With material, the cheap model spent 45 to 446 percent more prompt tokens per step in every family, and more requests and more seconds per step.
Little room to help
Without material, 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.
The small local model could not drive the harness
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.

Results by family

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.

What changed because of it

We stopped showing the phone item, and no demonstration on this site may use an item that a measurement found harmful.

What it does not show

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.

Limits and method

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.

Case study

Pi and Gemma 4 on Ollama Cloud.

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.

Checked end to end on September 23, 2026 · Pi 0.73.1 · Gemma 4 31B through Ollama Cloud

Get started Set up Pi

The three parts

The harness
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.
The model
Gemma 4 31B through Ollama Cloud. Pi calls the model with your own access; Baltor never asks for a model key.
The 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.

What happened

Found
The model searched the library with baltor_search. A search returns references, never file bodies.
Checked
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.
Installed
The skill landed in .pi/skills/<name>/SKILL.md. The extension never replaces a folder it did not install.
Loaded
A new Pi session loaded the skill. Pi lists a new skill from the next session on.
Used
The model used the skill in that session.

Try it yourself

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

Full steps on Get set up

What it does not show

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.

Case study

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.

Checked on the live service on September 24, 2026 · 11 of 11 steps passed

Get started

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.

Your address first. Get started asks for your email address and nothing else.
Baltor's own link. Baltor sends the confirmation link itself, from accounts@mail.baltor.ai.
Your password after the link. You choose your password on the page the link opens, so nobody sets it before you confirm.
One way in. An account exists only when Baltor's sign-up created it. Any other account under the same address is refused at sign-in, then archived and replaced when the owner of the address signs up through Baltor.

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.

The message is Baltor's
It came from accounts@mail.baltor.ai with the subject "Confirm your Baltor account", and its link opened the confirmation page.
The link signs the owner in
Following the link confirmed the address, set the password and signed the owner in.
A back-door account is replaced
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.

Pricing

One plan.
The whole library.

Baltor Pro is for one person and the devices that person connects. You bring your own harness and your own model access.

One plan

Baltor Pro

$29 a month

The whole library. Skills, instruction files, tools and code, subagents and commands, hooks and protocol server settings, searchable from your harness.
The measured unit is one downloaded item. One item at the version you selected counts once. A search, a second look at the same reference and a refused download are not downloads.
New vetted additions. New skills, tools and plugins join the library as they pass review, at no extra cost.
A key for each client you connect. Every client you connect gets its own key, issued to your account.
Set up for Claude Code, Codex, OpenCode, Pi and the Baltor Harness. Copy one reviewed entry. It never holds your key.

Get started Get set up

Subscribe from your account page once your account is ready.

Common questions

Do I need a model key to use Baltor?

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.

What happens to material I already downloaded?

It stays where you saved it. Revoking a key or cancelling stops new requests; it does not recall files that were already delivered.

Is there a free plan?

The Baltor Harness you install is free and open source. The library is Baltor Pro, $29 a month, and the measured unit is one downloaded item.

Can a team share one account?

Not today. One account holds one person's keys. Each person on a team needs their own Baltor Pro account.

How it works

Big problems.
Clear steps.

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

Prepare a customer import without changing the original.

Example only · no work runs here

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.

How this step could run

Information for this step

Selected files

What it can do

What it cannot access

What you get

When this step is finished

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

Give each step a useful starting point.

FindLook for relevant guidance and reusable code
ChooseCheck that it fits the task and your permissions
PrepareGive the step its selected files and instructions
WorkUse an appropriate model, tool or existing code
CheckTest the result against your requirements

A useful briefing is not always the shortest one. Start with relevant information and add more when the work reveals a gap.

What if a step needs more information?

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

Not every step needs a large model.

The goal and quality checks stay the same. The way a step does the work can vary.

Build something new

A focused AI workspace

Give your chosen development tool a clear brief, selected files and the permissions needed to build or repair the missing pieces.

Make a choice

A smaller, focused request

A decision may need one model request rather than a complete development session. Use a compatible model service you have already configured.

Reuse existing work

Run code that already fits

Use reviewed code when it meets the requirements. That step may need no model call and no newly generated code.

Models may run on your machine or through a provider you choose. You supply access; Baltor does not need to install or host the models for you.

Make the effort count

Is your AI spending effort in the wrong places?

These are the questions Baltor is designed to address at each step of the work.

Information

Too much context for a small task?

Match information to the task. Bring in more when it helps, without making every step carry the whole history.

Relevant information, less repetition

Model selection

An expensive model for every decision?

Compare suitable models and reusable code for each step, against your quality and spending requirements.

An approach that fits the work

Domain knowledge

Missing the right domain expertise?

Bring relevant methods, examples and checks into the work, with sources and applicability kept visible.

Specific guidance, not generic guesses

Code reuse

Paying to rewrite code that already exists?

Find an eligible implementation first. Generate new code for the gaps instead of reproducing a known solution as more output tokens.

Reuse what is already qualified

Improvement

Are the same mistakes showing up again?

Keep useful results and corrections. Compare proposed improvements on real tasks before making them the default.

Measured improvement, not blind repetition

Building blocks for better solutions

Put knowledge, code and experience to work.

Useful intelligence includes how to approach the work, code that can do it, what happened before, and what you want done differently.

01 / Approach

Context Intelligence

Methods, questions, constraints, examples and output contracts that guide the work.

For the import task

A field-definition guide and a checklist for missing values.

02 / Reuse

Code Intelligence

Reusable functions, tools, packages and workflows with declared inputs, outputs and effects.

For the import task

A qualified normalizer with its dependency and verification records.

03 / Experience

Runtime History and Solution Intelligence

Saved outcomes, failures, repairs, measurements and solution information that can inform a new task.

For the import task

A previous import's failed assumptions and the checks that detected them.

04 / Guidance

User Feedback Intelligence

Scoped corrections, priorities and instructions supplied by a person.

For the import task

“Keep uncertain matches for review. Never overwrite the original file.”

The four layers describe the broader product. The hosted service currently provides a small Context Intelligence example; it does not yet offer a populated catalogue across every layer.

Your tools, with help from Baltor

Get useful material. Keep control of your work.

Available today

With an account, you sign in with your email address and password, create a client token for each of your tools on your account page, search the library, download permitted files and see your usage.

The complete example above is still being tested from start to finish, and automatic setup of every step is not available yet. We have not established token savings or overnight task completion for this release.

Get started

Technical setup and architecture

Get set up

Three steps to your first download.

Create your account, connect the harness you already use, and let your agent ask for what each step needs.

Step 1 / Get access

1. Get access

Ask the operator for an account

The person who runs this service creates each account and issues its first key. To ask for one, write to the operator at the contact address in the privacy notice.

Step 2 / Connect your harness

2. Connect your harness

Your harness connects to this service with the Model Context Protocol. Pick it below and copy one reviewed entry. It holds this service's address and the name of the environment variable with your key, never the key.

Loading the reviewed connection settings Merge this entry into your existing configuration. Do not replace your other settings.

Loading the secret-free configuration.

Source: Client documentation

Loading documented compatibility information.

Set BALTOR_SERVICE_TOKEN in your environment. The file holds the variable's name, never the key, so it is safe to keep in version control. Then check the entry with Choose a client above.

Temporary setup in Bash

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 service token: " 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.

Check the service connection

Endpoint
Transport: Streamable HTTP
Protocol: Checking availability
Authentication: Service token, sent in the Authorization header

This browser check initializes the protocol and lists the available operations with your current connection. It does not run any work, fetch file bodies or spend model credits.

Not tested. No model calls are made by this check.

Configuration, connection, retrieval, native loading and useful task completion are separate checks. A green connection does not qualify all five. This service does not offer a browser-based authorization flow for native clients yet. Read the access and data boundaries.

When you no longer need it

Open your account page, find the token under Your client tokens and select Revoke. The service refuses a revoked token on its next request. Revoking a token does not recall files that were already downloaded.

Choose a client above to see how to remove its entry.

Open your account

Step 3 / Search and download

3. Search and download

Ask your agent to search Baltor for what the current step needs. It gets short references with source, licence and digest, downloads only the one it chooses, and every download appears in your usage.

Open the first example Search in the browser

Connecting needs nothing installed beside your harness. The engine that runs on your machine is free and open source under the MIT licence; the documentation has its installation steps.

Access and data

Know what you share.
Control what can run.

Connecting to the intelligence service grants access to permitted material. It does not give the service permission to run commands on your computer.

Service access and model access are separate

A service token identifies a scoped tenant connection. The service stores token 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 service token or client token entered on the sign-in page stays in page memory only, and reloading clears it. Avoid shared devices, untrusted extensions and screenshots of credentials.

Selected files do not grant execution authority

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.

How review works

An item joins the library only after independent reviewers approve its exact bytes. No reviewer judges an item it wrote, and one written rejection keeps an item out, with the reason recorded.

Since September 24, 2026, a new item 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. Each reviewer reads the item as a customer would and judges it 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.

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.

What reaches this service

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.

Do not send private customer content yet. The final retention policy, deletion workflow and consent controls remain launch work, and we would rather say that than let you assume otherwise. Embeddings and derived features must not be treated as anonymous data.

Current operational limits

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.

You sign in with your email address and password. Client tokens for your tools are created and revoked on your account page, and a client token cannot manage your account, your billing or other tokens.

Before sensitive or consequential work

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.

Get started

Privacy

Baltor privacy notice

Kind: published notice. The owner approved it on September 22, 2026, with Baltor.AI as the operator and the postal contact address below. It states what the hosted Baltor service stores today, taken from the source and the deployment. Planned behaviour is marked as planned.

Operator: Baltor.AI, 1428 Bryn Mawr St, Saxton, PA 16678, United States.

What Baltor is

Baltor serves reviewed material, such as guidance, skills and reusable code, to the tools that you run on your own computer. Your tasks, your files, your prompts and your model keys stay on your computer. The service never receives them and has no way to ask for them.

What the service stores

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 you your usage and, when billing opens, to reconcile it	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
A record of each refused request: a random reference, the address path, the reason code, the account when one is known, the time and the service version	To find and fix problems. The last 500 are kept and older ones are removed. A record never holds a password, a key, a promotion code, a request body or your network address.	The service database
Your payment details	Planned, when billing opens. Stripe collects and keeps them. Baltor receives only a customer identifier and the state of your subscription.	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.

What the service does not do

It sets no cookies. When you sign in with your email address, the page keeps your sign-in in your browser tab's session storage until you close the tab or sign out, so reloading the page keeps you signed in. A key you enter on the sign-in page is kept only in the memory of the page.
It runs no analytics, advertising or tracking scripts. Every script on the site is served from the site itself.
It keeps no log of requests that succeed. Refused requests are recorded as described above.
Email sent to you has open and click tracking switched off.
It does not sell or share your data. The providers named above process it only to run the service.

Backups and deletion

The hosting provider keeps daily snapshots of the service database for five days. To delete your account and its records, write to the contact address below. Deletion removes the account record, the key digests and the usage records; snapshots expire within five days. 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. Self-service deletion is planned.

Contact

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.

Changes

This notice changes when the service changes what it stores. The history of this file is the record of those changes.

Documentation

Set up, search and download.

Connect your own tools, find material for each task and check your downloads.

Get started First retrieval example Access and data boundaries

Loading the documentation list.

Run the engine on your machine

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.

Technical runtime reference

The public website calls a unit of work a step. The implementation uses the canonical Loop runtime described here and in the repository.

Runtime classification and role profiles

Complete behavior of a discrete cognitive or act step Loop node

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.

The perfect harness working directory for each task or subtask.

Start to finish, for three kinds of work.

Clean a customer file

Run a long job while you sleep

From the metric to the submission

Anything a harness can pick up, in one place.

Skills

Instruction files

Tools and code

Subagents and commands

Hooks

Protocol servers

Three things teams do with Baltor.

Solve complex problems overnight

More efficient operation

Learning and optimization, built in

Built for teams that check the details.

Your keys stay yours

Exact versions

Nothing runs on your computer

One plan. The whole library.

Baltor Pro

Questions, answered.

Give your agent the right setup for every task.

What teams do with Baltor.

Solve complex problems overnight

More efficient operation

Learning and optimization, built in

Solve complex problems overnight.

How a night of work runs

What you need

One harness per step, your choice

More efficient operation.

Only the context a step needs

Smaller models for smaller decisions

Reuse code instead of writing it again

The right tools, not every tool

Learning and optimization, built in.

Every run leaves a record

The next task starts smarter

Measured, not guessed

A library that keeps getting better

Your coding agent should not start from nothing.

Only what the step needs

Find the code before writing it again

Keep the large model for the hard part

Your machine stays yours

Built for the harness you run

One plan

Every agent on your team, working from the same expertise.

One reviewed library

Reviewed before it is served

Keys that belong to a person

One account for each person

What Baltor is, and what it costs.

The price

What counts

A record you can read

What Baltor is not

The protocol, the client and the setup.

The facts

References first, bodies second

Versions, on purpose

Ask the service first

One task, five small steps.

Profile the columns

Split the addresses

Find duplicates

Propose merges

Hand over the copy

Real searches

Your agent chooses

What a measured run found

A Kaggle competition, start to finish.

Fix the metric

Audit the splits

Train a baseline

Find where it fails

Check the score holds

Write the submission

One plan.
The whole library.