Baltor Get started

Harness and agent optimized operation

The right setup for every step your agent takes.

Baltor is a library of reviewed skills, instructions, tools and plugins. Your harness searches it, downloads only what a step needs, and checks every file before it lands in the working folder.

A harness is the program that runs a coding agent. Baltor sets up Claude Code, Codex, OpenCode, Pi and the Baltor Harness.

An example working folder for one step
working-folder/
├── AGENTS.md
├── .agents/skills/split-address-lines/
│   ├── SKILL.md
│   └── scripts/split_address_lines.py
└── .mcp.json

Each file is checked against its digest

The problem

Is your AI spending effort in the wrong places?

Coding harnesses and agent systems lose effort in the same ways, whatever the task.

  • Too much context for a small task

    Steps carry the whole history, so a small step gets slow and confused.

  • An expensive model for every decision

    Simple transformations go to the largest model because nothing picks a smaller one.

  • Missing domain expertise

    The agent guesses at methods a specialist already knows.

  • Rewriting code that already exists

    Tested code gets written again, and paid for again as output.

  • The same mistakes, again

    A correction made in one run does not reach the next one.

  • Long work small models cannot finish alone

    Large, multi-step problems stall without the right material at each step.

Source: the customer problems in AGENTS.md and README.md.

What Baltor is

One library for everything a harness reads.

Baltor puts the right skills, instructions, tools and plugins in each harness's working folder, so every step starts with the material it needs and nothing else.

  • Instruction files

    AGENTS.md, CLAUDE.md and other context files

  • Skills

    SKILL.md with its scripts, references and assets

  • Tools and reusable code

    Tested scripts and packages a step can run

  • Subagents and commands

    Agent and command definitions

  • Hooks

    Checks the harness runs by itself

  • Protocol servers and plugins

    The tools a step may call

Your models, your keys

Baltor never asks for a model key and never calls a model for you.

Reviewed before it is served

Every item starts as a candidate. Independent reviewers approve it before anyone can download it.

A fresh harness for each step

The Baltor Harness already runs a task on your model in small steps. Starting a fresh standard harness for each step is not open yet.

Source: README.md, what Baltor gives your harness, one harness for each step, and what is not open yet; the harness file definition in AGENTS.md.

How it works

Search returns references. A step downloads only what it chose.

  1. Search

    The harness asks the library for what a step needs. The answer is a short list of references: name, purpose, source, licence, exact version and digest. No file bodies.

  2. Choose

    The step picks one reference. An item that needs file, network or process access is offered only to a step that declares that access.

  3. Download and check

    Access is checked again at download. The file must be exactly the version the reviewers approved, and the download is recorded in your usage.

  4. Place

    The file goes where the harness reads it: a skill folder, AGENTS.md or a protocol server entry.

Source: README.md, searching and retrieving and usage and what you pay for.

One step getting one file from the Baltor library The harness searches the library and gets back references only. The step chooses one and downloads it. The library checks access again and returns the exact bytes. The harness checks the bytes against the digest and places the file in the step's working folder. Your harness at one step Baltor library reviewed items search the library references only name · licence · version · digest chooses one download that item access checked again the exact bytes recorded in your usage exact reviewed version placed where the harness reads it: .agents/skills/split-address-lines/SKILL.md

Live today

Running at baltor.ai, with every check on the record.

Release 24

Deployed on September 24, 2026 by image digest, from a commit whose continuous integration run passed. Release 23 is kept for rollback.

Source: the release 24 record

Public sign-up

Open since September 24, 2026: your email address, then the link we send, then a password you choose. No other way in.

Source: MVP-CLIENT-SERVER.md, the current deployment

5 harnesses

Connection entries for Claude Code, Codex, OpenCode, Pi and the Baltor Harness. Pi, running Gemma 4 on Ollama Cloud, found, installed and used a skill end to end on September 23, 2026.

Source: client-recipes.json, the reviewed connection entries

43 reviewed skills

Each approved by three independent reviewers, none of them its author. One written objection withholds approval.

Source: the release 24 record and reviews.json

Licence and digest

Every served item names its licence and its digest. After the release, the catalogue check compared the published digests with the served bytes: 7 of 7 passed.

Source: the release 24 record and the served manifest.json

Checked live

Browser checks, each hostname
163 of 163
Public addresses answering
120 of 120
Service checks
19 of 19
Sign-up journeys
11 of 11
Staff dashboard
10 of 12

Both staff dashboard failures were mistakes in the checking script; the failed runs are kept.

Source: the release 24 record, its live checks

Measured in the open

We test our own library, and publish the result when it says no.

A data cleanup study compared library items with no item, on a design frozen before the first model call and scored by an independent checker. Each step ran in a fresh, isolated harness process that held only its own material.

225 of 400

Requests used of the declared ceiling, across 72 counted steps. Every request reported its token usage.

Source: REPORT-2026-09-22.md

0.842 against 0.991

An approved phone number item made a cheap model clearly worse: 0.842 in all three repetitions, against a mean of 0.991 without it. The item's method was incomplete, and the report names the rows it lost.

Source: REPORT-2026-09-22.md

0 of 4

Task families where the material was clearly better; it was clearly worse in 1. The material cost 45 to 446 percent more prompt tokens per step.

Source: REPORT-2026-09-22.md

Next: a harder population, where the model fails without help, reusing the same frozen design, scorer and isolation recipe. The report asks for the measured harm to go into the item's review record; a revised item needs a new review and a new measurement.

The library plan

Where the library goes, and the rates measured so far.

Only the first step is live. The rest is the plan, and a candidate counts only after an independent review approves it.

  1. Live

    43

    reviewed skills in release 24

    Source: the release 24 record

  2. Plan

    10,000

    approved packages, the first milestone

    Source: the decisions in AGENTS.md

  3. Plan

    100,000

    approved packages, then 100 to 1,000 more each day

    Source: roadmap step S-6.69 in roadmap.yaml

  4. Aim

    Millions

    of licence-cleared files, continuously updated

    Source: MILLION-HARNESS-SUPPLY-NEXT-STEPS and S-6.40 in roadmap.yaml

Measured so far

  • 43 of 49 candidates approved by all three reviewers in the first review round.

    Source: reviews.json

  • 0 of 30 approved in the first pass of the multi-family review panel, each rejection written down with its reasons. The whole run, calibration included, took 658.5 seconds.

    Source: the first review panel run

  • 56 of 75 original multi-file packages passed the deterministic pre-checks in wave 5.

    Source: the wave 5 record

  • 3,251 outside candidates staged for review in one import run, from 4,790 discovered; 187 kept as outlines only, because their licence permits no copy.

    Source: the ingestion record

Business model

One plan. The whole library.

Customers bring their own model access, so model costs stay with their own provider. Baltor sells reviewed material and its delivery, never model calls.

Baltor Pro

$29 a month

  • The whole library, searchable from your harness
  • One downloaded item is the measured unit, with no overage billing
  • A personal key for each client you connect
  • Cancel any time from your account page
Get started

Source: README.md, status on the day of release 24

Why now

The formats became standards. Choosing and trusting the right file did not.

33.9 to 50.5 percent

Curated skills raised the mean pass rate from 33.9 to 50.5 percent on 87 tasks in SkillsBench, June 2026. Skills an agent wrote for itself scored 8.1 to 11.5 percentage points below no skills at all.

Sources: SkillsBench on arXiv, as recorded in the first-party package research

157 of 98,380

Skills confirmed malicious in a February 2026 study of 98,380 skills from two registries. Supply is plentiful; a reason to trust a file is not.

Sources: the study on arXiv, as recorded in the harness and library plan

Baltor's answer: a reviewed library, delivered one step at a time, with the licence and the exact reviewed version checked before a file is used.

Contact

Talk to us.

Create an account at baltor.ai and connect your harness, or write to the operator at the postal address in the privacy notice.

Baltor.AI 1428 Bryn Mawr St Saxton, PA 16678 United States Source: the privacy notice, PRIVACY-NOTICE.md
Get started Get set up baltor.ai