Autopus-ADK CLI
CLI installs a verify-by-running harness into Claude Code and Codex: build and tests, diff review and a separate QA agent.
- 109 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The hook runs builds, tests and your product with user permissions and can send the agent back to edit.
- `auto secure` runs attacks against the target you specify, and the installer runs from `curl`.
Install
- Works in
Manual install
curl -sSfL https://raw.githubusercontent.com/autopus-ai/autopus-adk/main/install.sh | shWindows has `install.ps1`. The installer does not run `auto init`.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The hook runs builds, tests and your product with user permissions and can send the agent back to edit.
- `auto secure` runs attacks against the target you specify, and the installer runs from `curl`.
This is third-party code. Review the repository files before installing.
What it does
Autopus-ADK adds a verification loop to an agent CLI (Claude Code, Codex, Antigravity CLI, OpenCode, Oh My Pi). The agent still plans and writes the code, while Autopus builds the project, runs tests, reviews the diff and has a separate QA agent run the product and hunt for bugs before the session ends. `auto init` registers a Stop hook in Claude Code: `auto verify stop` builds and tests only the changed toolchains and sends the agent back if the build fails. `/auto review` runs a reviewer on `git diff HEAD`, `/auto qa` invokes the `qa-explorer` explorer and `auto qa hunt verify` re-runs each reproduction. `auto secure scan` also runs the scanners you already have and prints one summary.
Who it is for. Developers who trust an agent to write code but will not accept work without a build, tests and an independent check.
Good fit when
- The agent reported it was done, but the project does not build or tests were never run.
- Before merging you want a diff review by a separate agent and a script that re-runs found bugs.
- A QA agent has to start the service and hunt for failures in real behavior.
- One command has to combine results from gitleaks, npm audit and other scanners.
Not a fit when
- You need a full planning pipeline with SPEC documents and dozens of agents; it was removed after version 0.50.
- The product and scanners cannot run on the machine where the agent works.
- Automatic fix rounds spend budget and you need control without repeated runs.
Example request
Run the change check with `auto verify stop`, then `/auto review` and `/auto qa`, and show the confirmed bugs.Limitations
Verification runs the build, tests and the product itself with your permissions, so read a reproduction before running it by hand. Review and QA cost noticeably more than build and tests, so they run on command or after opting into the follow-up round. `auto secure` attacks only your local or staging target and reports only confirmed findings. The author notes that the pilot comparison on four tasks shows a trend rather than a benchmark. The installer runs through `curl`, and installation does not run `auto init`.
How to disable. Delete the `auto` binary and the project files created by `auto init`, then remove the Stop hook from Claude Code settings.
FAQ
How do I install Autopus-ADK?
Run the one-liner `curl -sSfL https://raw.githubusercontent.com/autopus-ai/autopus-adk/main/install.sh | sh`. Alternatives: Homebrew `brew install --cask Insajin/autopus/auto` and `go install`.
What does auto init do?
It sets up the project interactively: it registers hooks and files for the chosen platforms. `--platforms claude-code,codex` picks platforms and `--yes` accepts defaults.
When does the Stop hook run?
Every time a Claude Code session tries to stop with a changed tree. The hook builds and tests only the touched toolchains and sends the agent back to fix failures.
How do I update Autopus-ADK?
`auto update --self` updates the binary with publisher signature and SHA256 verification. `auto update` regenerates project files from the installed version.
What does auto secure scan check?
It runs the scanners you have installed, such as gitleaks, govulncheck, `npm audit`, `cargo audit` and pip-audit, and prints one summary. `--staged` and `--fail-on critical` narrow scope and threshold.
Discussion
No questions or reviews yet.
Related
CLI toolsA CLI that finds traces coding agents leave in code, like filler comments, swallowed errors and fake imports, and scores it 0–100
Agent Reliability Analyzer statically scans AI agent code, finds reliability gaps with 340 plus rules and applies fixes
Reviews the git diff and the conversation with the agent: catches bugs and gaps between your request and the agent's result. CLI, skill, GitHub Action
A local debugger for TypeScript agents: run step trees, CI trajectory checks and portable evidence bundles, with no account.
Hooks for Claude Code and Codex CLI keep a ledger of agent actions and refuse a "done" claim until a check has passed.
OpenQodex reviews your change before a push: scanners with no model first, then a separate Claude Code or Codex process triages findings on your login
README in short
Modern models plan and write code themselves but check results poorly, so the project keeps one loop: run the change and send the agent back if it does not hold.