Ouijit
A git worktree-based task and terminal manager for running Claude Code, Codex, Pi and OpenCode coding agents in parallel.
- Source repo
- ouijit/ouijit
- Stars
- ★ 184
- Last updated
- today
- License
- AGPL-3.0
- Primary language
- TypeScript
- FA score
- 76/100 · Good
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A solo developer working on several features at once: each card gets its own worktree so multiple agents work on non-conflicting branches in parallel.
- Not a fit if
- Windows-only developers
- Users unwilling to install and sign in to an agent CLI locally
- Teams needing a hosted board or multi-device sync (data stays in local SQLite)
- Website
- ouijit.com ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-21
- Source review
- 76/100 · Good
What does this agent do, and when should you use it?
Ouijit is a free, open-source (AGPL-3.0) desktop app for running multiple coding agents in parallel. Each task card gets its own isolated git worktree and integrated terminal, and lifecycle hooks launch the agent with the task's prompt when the task starts. A built-in ouijit CLI is on PATH in every terminal, letting agents create tasks, move board cards and open panels beside themselves. Tasks live on a kanban board with To Do, In Progress, In Review and Done columns, and diffs, pull requests and issues are reviewed in the same window — a note on a diff line goes to the agent working in that worktree. Everything is stored locally in SQLite with no account, no sign-in and no telemetry, and agents can run sandboxed.
Tasks are managed as kanban cards; dragging a card between columns fires the matching lifecycle hook, and starting a task creates an isolated git worktree — a copy-on-write clone with node_modules and other gitignored files intact, or a clean checkout — with per-task merge targets and parent-branch task chains. Each task gets a terminal in a stack, with panels attaching as tabs: a runner for the dev server, a web preview, and markdown files with Mermaid diagrams. Inside Ouijit terminals, claude, codex, pi and opencode resolve to wrappers around the real binaries that add status hooks and the ouijit CLI reference. CLI subcommands such as ouijit task create, ouijit task set-status 5 in_review, ouijit hook set start, ouijit pr draft add, ouijit markdown add and ouijit preview add let agents and users drive the app from any terminal, with JSON output. Every task terminal includes a word-level-highlighted diff of the worktree against its merge target; line notes re-anchor as code moves and reach the working agent. An experimental GitHub surface drives the gh CLI for a PR inbox, locally staged review comments and merging, while the issues list is fed by GitHub and/or Linear. A ⌘K command palette jumps between terminals, projects, tasks and pull requests. Sandboxing is provided by nono under Seatbelt/Landlock, per project profile file or per command via ouijit sandbox -- claude …. Quitting saves the session for restoration on next launch.
- A solo developer working on several features at once: each card gets its own worktree so multiple agents work on non-conflicting branches in parallel.
- A reviewer who wants agent fixes: leave a note on a diff line and it reaches the agent working in that worktree.
- A maintainer turning issues into work: pull an issues list from GitHub or Linear and start a task on the suggested branch name.
- A security-conscious user: run agents in a nono sandbox while the dev server beside them stays on the host.
- A user who wants agents to self-organize: agents file follow-up tasks, move their own card and open a plan panel next to themselves via the CLI.
How do you install or deploy this agent?
Download the latest self-contained release for your platform:
- macOS (Apple Silicon): https://github.com/ouijit/ouijit/releases/latest/download/ouijit-darwin-arm64.dmg
- macOS (Intel): https://github.com/ouijit/ouijit/releases/latest/download/ouijit-darwin-x64.dmg
- Linux (x64): https://github.com/ouijit/ouijit/releases/latest/download/ouijit-linux-x64.zipRequirements: macOS 13+ (Apple Silicon or Intel) or Linux x64, git 2.20+ on PATH, and at least one installed and signed-in agent CLI (claude, codex, pi or opencode). The PR and issues surfaces additionally need gh 2.48+ with gh auth login.
Building from source requires Node.js 22.22+ and C/C++ build tools for node-pty (macOS: xcode-select --install; Debian/Ubuntu: sudo apt install build-essential python3):
git clone https://github.com/ouijit/ouijit.git
cd ouijit
npm install
npm startHow do you use this agent?
Launch Ouijit, add a project and create tasks on the kanban board. Starting a task creates the worktree and opens a terminal; configure a hook to launch the agent with each task:
ouijit hook set start --name "Agent" --command 'claude "$OUIJIT_TASK_DESCRIPTION"'Then use the ouijit CLI in any Ouijit terminal (or your own shell, with JSON output):
ouijit task create "Fix flaky login test"
ouijit task set-status 5 in_review
ouijit pr draft add 116 --file src/api.ts --line 88 --body "throws when the token is missing"
ouijit theme use dracula
ouijit markdown add ./plan.md
ouijit preview add http://localhost:3000Use ⌘K (Ctrl+K on Linux) to jump between terminals, projects, tasks and PRs; write notes on diff lines to hand them to the agent in that worktree; run agents sandboxed with ouijit sandbox -- claude …. Quitting saves the session and the next launch offers to restore terminals and panels.
What are this agent's strengths and limitations?
- Every task automatically gets an isolated git worktree, so multiple agents can work in parallel without clobbering each other's branches.
- The
ouijitCLI lets agents create tasks, move cards and open panels with no setup, shaping the app from within a session. - Diff-line notes go straight to the agent in that worktree and re-anchor as code moves, closing the review loop in one window.
- Fully local storage (SQLite) with no account and no telemetry, plus optional sandboxed agent execution via nono.
- Supports four harnesses — Claude Code, Codex, Pi, OpenCode — without changing anything outside Ouijit's terminals.
- Only macOS 13+ and Linux x64 builds are published; there is no Windows release.
- You must install and sign in to at least one supported agent CLI, and bear the associated model/subscription costs.
- Pull requests are not accepted — the project is read/build/fork only; bugs and feature requests go through the issue tracker.
- The PR inbox and merging depend on the
ghCLI and are marked experimental; issues integration is limited to GitHub and Linear. - The GitHub surface requires gh 2.48+ plus a
gh auth loginstep — extra configuration on top of the app itself.
How does this agent compare with similar options?
The README explicitly names Claude Code, Codex, Pi and OpenCode as supported agent harnesses; Ouijit does not compete with them but orchestrates them in parallel with worktree isolation and kanban management. Support for any other harness must be requested via the issue tracker.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Ouijit This agent | 76 · Good | Desktop appFree + model costs | ★ 184 | today | TypeScript | Codex · Claude Code |
| Prometeu | 73 · Some gaps | Desktop appFree + model costs | ★ 67 | 1d ago | Rust | Codex · Claude Code |
| Skiff | 72 · Some gaps | Desktop appFree + model costs | ★ 21 | today | Rust | Codex · Claude Code |
| Agentbox | 66 · Some gaps | Desktop appFree + model costs | ★ 451 | today | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows a nono-based sandbox (Seatbelt/Landlock) selectable per terminal/hook/command, workflow permissions scoped to contents: read, and PR review comments staged locally until sent — supporting least-privilege and external-effects. Deductions: the local REST API lets agents create tasks, move cards, and switch themes with no documented mandatory confirmation gate; the no-telemetry claim is a README assertion with no code shown; no documented undo/recovery for merges or task deletion; gh/Linear credentials are delegated to external CLIs without handling evidence in the packet.
Evidence shows strong internal consistency: README, AGENTS.md, package.json scripts, and CI workflows agree (docs:check keeps CLI docs synced; checks.yml explains why E2E is excluded). CLI maps ApiError to clean stderr plus non-zero exit. Deductions: postinstall downloads the nono binary over the network at install time; integration tests are skipped in CI due to missing git identity (#63), so coverage is incomplete.
Evidence shows a sharply defined audience (developers running Claude Code/Codex/Pi/OpenCode in parallel) and thorough environment fit (macOS 13+/Linux x64, git 2.20+, gh 2.48+, Node 22.22+, build tools documented). Deductions: agent capability boundaries are not systematically enumerated (which CLI actions an agent may take autonomously vs. with confirmation is not partitioned); trigger precision is evidenced well only by the single sampled SKILL.md.
Evidence shows the full AGPL-3.0 text matching the README, complete install notes (binaries plus source build), and clear information architecture. Deductions: package.json's description ('Discover and manage your local development projects') contradicts the README's framing; FAQ and full docs live only on the external site and are absent from the packet; no CHANGELOG file (only a version number and auto-update notes); maintenance is issues-only with PRs explicitly rejected, and the author (pbjer) is unverified.
Evidence shows good output usability (uniform JSON stdout, stderr errors, non-zero exits) and genuine marginal value via per-task git worktrees with integrated agent terminals. Deductions: cost/benefit is only statically inferable — installing downloads an externally built nono binary and requires several external agent CLIs, and the E2E suite's runner behaviour is admitted by the project itself to be unknown, a known unverified cost.
Evidence shows strong cross-source corroboration: README claims (local SQLite, gh-driven PR surface, CLI commands) match package.json, source, and workflows, and docs:check enforces doc/command sync. Deductions: claims like no-telemetry and copy-on-write cloning cannot be traced to implementation within the provided files, and documents do not separate stated facts from inferences.
- The local REST API lets agents create tasks, move cards, and switch themes autonomously with no documented mandatory confirmation gate; evaluate restricting it before enterprise deployment.
- postinstall downloads the nono sandbox binary over the network; audit this supply-chain dependency before production use.
- The no-telemetry claim is a README assertion without implementation evidence in the packet; verify via network traffic for sensitive environments.
- No documented recovery path for merges or task deletion; keep your own branch backups.
- package.json's description contradicts the README and there is no CHANGELOG; rely on external release notes when assessing behaviour changes between versions.
FAQ
Does Ouijit itself cost anything?
Do agents get access to my whole machine?
ouijit sandbox -- claude … sandboxes a single command — the agent runs contained while the dev server beside it stays on the host.Where is my data stored?
Which operating systems and agents are supported?
Can I contribute code?
Is Ouijit open source?
Is Ouijit safe to use?
Is Ouijit still maintained?
Who makes Ouijit?
What is Ouijit's official website?
Thanks — we use this to improve the page.