Agetor turns a kanban board into a local AI agent controller

Agetor is an open-source orchestrator for CLI coding agents that runs entirely on the user’s machine. According to the project site, it is designed for Claude Code, OpenAI Codex, and similar tools, and it presents them through a kanban board that manages many tasks and many repositories at once.

The app treats each kanban card as a task made up of a prompt, a working directory, and an agent choice. When a user starts a task, Agetor spawns the selected agent as a child process, isolates it in a git worktree, streams output back to the interface, and moves the card through columns as the run progresses.

The project says this setup is local-first. There is no cloud relay and no remote sandbox, and agents run with the same shell privileges they would have if launched directly by the user in a terminal.

Agetor also supports multiple agents and multiple accounts. Built-in support exists for Claude Code and Codex, and users can define additional harnesses to run a second Claude or Codex account in parallel. Each harness gets its own dedicated $HOME directory so logins, history, and configuration do not collide.

For task isolation, Agetor creates a per-task git worktree under ~/.agetor/worktrees/ and pins the base ref at creation time. The project says that lets two agents work in the same repository at the same time without stepping on each other, and repeat runs can start from the same commit for consistency.

Claude Code tasks run in a per-task tmux session that stays alive across multiple turns. Agetor reads Claude’s JSONL transcript and renders assistant text, thinking blocks, tool calls, and tool results in the UI. Codex runs are handled differently: each one is a fresh codex exec invocation, and follow-up messages create a new run record on the same task.

Approvals and questions are also handled in the interface rather than buried in a terminal view. According to the project, a PreToolUse hook and a small MCP server intercept Claude’s permission prompts and ask_user calls, then surface them as structured cards with radio buttons, checkboxes, and text fields. Users can approve, deny, or allow a tool for that task only.

Agetor keeps a live event stream for each task and a global channel for status updates. It uses SSE, or Server-Sent Events, to push run updates, and it sends native notifications when a run finishes, succeeds, fails, or blocks waiting for input.

State lives in a local SQLite database under ~/.agetor/agetor.sqlite. The project says the database stores tasks, runs, events, projects, harnesses, preferences, and approval rules, and nothing leaves the machine.

The app itself is built as an Electrobun desktop app, with a Bun main process and a native webview for the UI. Agetor exposes a localhost HTTP API protected by a per-launch random bearer token, and the project says that token is passed into the webview via preload rather than appearing in the URL.

The project is targeted at macOS first, with Linux and Windows builds configured but not yet tested. It requires Bun, tmux, Git, and at least one agent CLI on PATH, and the repository includes install and development instructions for running it locally.