Agentplane CLI
Agentplane handles task state, branches and result evidence in a CLI while the coding agent focuses on the content of the change.
- 86 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The tool drives Git and pull requests for tasks by running commands in the repository.
- Managed `agentplane task run` executes episodes without launching the agent by hand.
Install
- Works in
Manual install
npm install -g agentplaneRequires Node.js 24 or newer, Git and a terminal. Then run `agentplane init` at the project root.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The tool drives Git and pull requests for tasks by running commands in the repository.
- Managed `agentplane task run` executes episodes without launching the agent by hand.
This is third-party code. Review the repository files before installing.
What it does
Agentplane splits work between a person, a coding agent and a command-line tool. The person sets the outcome and accepts risk, the agent understands the problem and writes code and tests, and the tool stores task state, routes Git and pull requests, collects evidence and closes the work. The `agentplane task advance` command hands the agent one bounded episode: objective, writable scope, context, result schema and the exact command to return it. The agent cannot change the task lifecycle itself. The tool rejects stale results by state fingerprint and builds evidence from observed repository changes and check results.
Who it is for. Teams where coding agents edit a shared repository and need a verifiable trail of who changed what and what confirms it.
Good fit when
- An agent closes a repository task and the reviewer needs facts about edits and check results, not the agent word.
- Codex, Claude Code and Cursor work on one project and all need the same task contract.
- Changes must go through an approved plan and stop for human confirmation.
- After an agent failure the task must resume without losing state and check records.
Not a fit when
- The repository is small and personal, and a process of plans, tasks and confirmations is overkill.
- The tool does not make code correct: a person or the agent still judges the quality of the solution.
- The machine has no Node.js 24 or newer.
Example request
Create a task through Agentplane, perform the issued episode and return the typed result with the command the tool printed.Limitations
Node.js 24 or newer, Git and a local terminal are required. The tool makes the process around the model deterministic, not the model itself: semantic decisions stay with the agent. For an existing repository `agentplane init` runs from the root: it creates a policy gateway and a local work contract, and needs no account. Codex, Claude Code, Cursor and Aider all fit. Managed runs with `agentplane task run` work only after an approved plan and a configured runner.
How to disable. Remove the global package with `npm uninstall -g agentplane` and delete the `.agentplane` folder in the repository once the policy is no longer needed.
FAQ
How do I install Agentplane?
Run `npm install -g agentplane`, then `agentplane init` at the repository root and `agentplane quickstart`. No account is needed.
Which agents work with Agentplane?
The episode from `task advance` suits Codex, Claude Code, Cursor, Aider or any other agent that can edit a repository.
What does the agent get from task advance?
A packet with the objective, writable scope, context, result schema, the `exchange.result_path` path and the `exchange.resume_argv` command for returning.
What is the ap command?
The short alias `ap` replaces `agentplane` in calls made by the agent.
When does the tool stop and wait for a person?
At approval, external-effect, recovery and terminal boundaries. After that the agent continues only through the next command the tool prints.
Discussion
No questions or reviews yet.
Related
CLI toolsAgentic OS adds a five-step workflow and checks in git hooks and CI so that a step without evidence does not count as done
A plugin marketplace and installer for the full dev loop: task selection, review, cleanup, PRs and releases
A CLI and skill that record an agent's goals, decisions and rejected paths in Git so the next agent does not repeat them
Orchestrator skill: from an idea the agent writes requirements and a spec, splits work into tasks, builds the project, checks it
One desktop window for Claude Code, Codex, Gemini CLI and OpenCode: terminals, pull requests, diffs and approvals
APM
APM (Agent Package Manager)
A package manager for AI agent context: skills, prompts and MCP servers install with one command across supported agents
README in short
The tool owns rules, state and evidence, the agent owns the meaning of edits. Configuration lives in the `.agentplane` folder of the repository.