no_human
github.com/no-human-ai/no_human- Category
- AI Agents
- Rank
- No. 1139Tools index
Previous survey · No. 1127 ·
- Pricing
- Open Source
- Type
- AGENT
- Use case
- Coding
- GitHub
- 330 stars
- Latest release
- v0.2.4
- Date
About
no_human is a free, open-source tool that automates software development from an initial ticket through to a reviewed pull request, running locally on the developer's own machine rather than as a hosted service.
What it does
A work queue for AI coding agents that you host yourself. You hand it an issue link or pull tickets from Jira, Linear or monday.com; it writes a plan with acceptance criteria, has a Claude agent make the change in a separate branch, counts any deleted or weakened tests, then sends the diff to a second model that never saw the coder's session and is told to refute that the work is done. If the checklist passes, it opens a pull request and stops. A person approves the merge from a local web board, a terminal command or a desktop app.
Why it's ranked here
MIT licensed, installable from PyPI, with a command line, a local board, an MCP server mode and a GitHub Action built on one reviewer. What sets it apart is how much the documentation admits. The security guide states plainly that the tool is not offline and that the coding agent runs with every tool call pre-approved by default. A known-issues file publishes measured failure rates and what a fix had to prove. The project reports that over 65 days of building itself, its reviewer sent back 505 of 1,709 attempts the coder called done. That figure is self-reported, but the gates behind it are described in checkable detail.
What's good
The agent cannot merge. Landing a change is a human command that squashes the branch and pushes under your identity, and agent sessions carry a marker that the command line and board refuse. Bug fixes can be held to a reproduction rule: the offered test must fail on the old code and pass on the new one, which stops a fix that proves nothing. Billing is pinned to one path: in the default subscription mode a stray Anthropic API key aborts startup instead of silently charging metered usage. The packaging notes record past shipping mistakes, such as wheels that once carried no database schema, and the guard added for each.
Tradeoffs
By default the coding agent has a shell and nothing sits between it and the network; the docs say the only real control is a container or egress proxy. Every task pushes a branch to your git host, and no setting disables that. Usage telemetry is opt-out, on unless you turn it off. The default coder and reviewer are both Claude models on your own credential; other backends swap only the coder. Parallel workers ship off, and a database crash under concurrency was closed only after measurement the project itself bounds near one percent. Reviews are slow: a measured round took about 18 minutes. The manifest version is pre-1.0.
How to use it well
Fits a small team that already tracks work in Jira, Linear or GitHub issues and wants agent-written pull requests that arrive with a plan, test results and a reviewer checklist citing file and line. Run it inside a container or behind an egress proxy, keep one worker until you trust parallel runs, and switch the reproduction gate to required so every change must prove itself. If you only want the second-opinion reviewer, the GitHub Action judges a pull request's diff alone, capped at 15 files by default. It does not help in an air-gapped setup or without an Anthropic credential.
Technical notes+
pyproject.toml declares no-human with a Python >=3.12 floor, MIT via PEP 639 license-files, hatchling build, console scripts nh and no-human both pointing at the click entry point, and an mcp>=2,<3 bound whose comment ties both bounds to the MCP bridge import after an unbounded pin broke installs. The wheel and sdist exclude the internal ci_gate package and a private vendor-terms file, and force-include the SQL migrations after shipped wheels once created no schema. src/no_human/api/app.py is a FastAPI app that resolves the built board from either the checkout or the installed wheel, starts the embedded scheduler in its lifespan, and records an app_started telemetry event (comment: opt-out, default on). docs/security.md documents bypassPermissions as the coder default, acceptEdits as the narrower option, environment scrubbing of competing Anthropic, Bedrock and Vertex credentials, and a session mark it calls a tripwire, not a lock. docs/configuration.md lists claude-sonnet-5 as implementer, claude-opus-4-8 as reviewer, 1500s review timeout, and single-turn review for diffs under 200 lines unless a risk-flagged path is touched. action.yml runs a Docker image with max_files 15 and skips fork PRs. src/no_human/brain/cli.py adds an off-by-default team rule sync with PKCE login on loopback port 8422 and signed manifests. docs/KNOWN_ISSUES.md records KI-1 closed after 300 clean serial runs.
Observed
- License
- MIT (SPDX expression in the package manifest)
- Language
- Python 3.12+, with a React web board built by npm
- Install surface
- PyPI package installed via uv or pipx; desktop builds for macOS, Windows and Linux; source install via uv
- Interfaces
- CLI, local web board on loopback, MCP server mode, GitHub Action (Docker)
- Model backend
- Claude Agent SDK on the operator's own credential; codex and local backends for the coder only
- Storage
- Local SQLite database; no Postgres or Redis
- Merge policy
- Agent opens the PR and stops; merging is a human command
- Tracker intake
- Jira Cloud, Linear and monday.com polled; GitHub and GitLab issues by URL
- Telemetry
- Opt-out, enabled by default
Read from README.md, pyproject.toml, docs/README.md, docs/security.md, docs/configuration.md, docs/KNOWN_ISSUES.md, src/no_human/api/app.py, src/no_human/brain/cli.py, action.yml.
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.