Automatic Control for intelligence in action.
Quickstart · Documentation · How it works
Tell an agent what to do in your own words. A language model plans and writes short programs; The Harness runs them in an environment the agent can see and act in. Watch, send a message, or take over at any time.
The agent thinks while it acts: between model calls, its programs, skills and reactive rules keep working. What it learns along the way—notes, tested skills and rules—applies in the same run.
npm install -g @lmmslab/harness
harnessThe Harness needs Python 3.11 or newer. You can also install it with pip install the-harness.
harness opens a conversation in your terminal. The first time, it asks which model to use and for its key; then type a task. Any model works: hosted ones through OpenRouter, OpenAI, Anthropic and other providers, or a model you run yourself.
See Install and the Quickstart for a real task on your own computer.
Use The Harness in the terminal, run one task from the CLI, drive it from another program with JSON lines, call it from apps, editors and other agents over MCP and ACP, or build on the Python API.
harness # talk with the agent in the terminal
harness run "task" -p # do one task and print the answer (--json: every event)
harness run rpc # drive a conversation with JSON lines, from another program
harness install # packages (also: list, remove)
harness library # what the agent learned
harness replay # a run again in the terminal
harness doctor # what this computer has and lackspip/npm install only the small core. Environments come in packages, installed by themselves the first time a task needs them.
| Environment | Package |
|---|---|
| Your computer: apps and windows, or a display of its own | computer |
| Websites, in a background browser | browser |
| Project files and processes, in a bound working directory | workspace |
| Package | What it is | Depends on | Provides |
|---|---|---|---|
| acp | The Harness as an Agent Client Protocol agent for editors such as Zed | – | commands acp |
| agent | Persistent identity, multi-turn dialogue and private memory over the Harness task engine | – | commands agent |
| browser | Search and browse the web in a background browser | computer | skills browser; benches browser |
| computer | Control a computer: its desktop, its applications in the background, displays of its own | – | embodiments computer, ubuntu; skills desktop, physical, screens, windowtools; benches computer |
| eval | Evaluate the agent: experiments over tasks and seeds, their suites, locks, run pages and self-improvement | local-models | commands config, inspect, lock, program, runs, self-improve, serve, view |
| external-agents | ACP, terminal and cloud coding sessions, and bounded Claude Code and Codex delegates | – | commands agents |
| local-models | Models on your own GPUs: where services run (deploys), their weights and servers, Qwen3-VL | – | services qwen3-vl-32b, qwen3-vl-8b; commands down, fetch, services, status, up |
| mcp-example | An example MCP server: unit conversion | – | services mcp-units |
| mcp-server | The Harness as an MCP server (harness mcp), for apps, editors and agents that are MCP clients | – | commands mcp |
| perception | See: segmentation, pointing, scenes, generic vision and vision-language models | local-models | skills pointing, scene, segmentation, vision, vlm; services molmo2, sam3, vision |
| workspace | Operate project files and processes in a bound working directory | – | embodiments workspace; skills workspace; benches workspace |
- Quickstart
- Automatic Control: the ideas behind The Harness
- Models and providers
- Interrupt and take over
- Let it learn an environment
- Write a skill, write a package, add an environment
pip install -e .
harness install ./packages/computer ./packages/browser ./packages/workspace --link --yes
harness doctorSee CONTRIBUTING.md and the validation guide.
@misc{lmmslab2026harness,
author = {{LMMs-Lab}},
title = {The Harness: Automatic Control for Intelligence in Action.},
howpublished = {\url{https://github.com/EvolvingLMMs-Lab/the-harness}},
year = {2026}
}MIT; third-party material retains its own terms. See LICENSE and CONTRIBUTORS.md.