Pleiad is a desktop app for working with several coding agents side by side.
Conversations, statuses, branches, and approvals work the same way no matter which agent is behind them.
Features
- One list for every agent: Claude Code, OpenAI Codex, Google Antigravity, OpenCode, Devin CLI, and Grok Build conversations in a single list, organized by status and group.
- Compatible endpoints: Point Claude Code or Codex at OpenRouter, LiteLLM, a local Ollama, and more, and pick one per conversation.
- Branching: Fork a conversation from any message to try another approach.
- Approvals: Approve tool calls right in the conversation, or set a mode per conversation, from asking every time to fully automatic.
- Delegation: Agents can hand work to other agents. Pleiad picks the target from the kind of task and remaining usage.
- Stay in the conversation: Preview files and open web pages in a side panel, and view HTML that agents render in the chat.
[!NOTE]
Pleiad is in beta. Features, UI, and the stored data format may change without notice.
Install
Download from pleiad.dev/download or Releases.
- Windows 10 / 11 (x64 or ARM64). Beta builds are not signed by a trusted publisher, so Windows SmartScreen shows a warning on first launch. Choose More info → Run anyway.
- macOS 13 or later (Apple Silicon or Intel). Signed and notarized.
- Linux: run the browser version from source (see below).
- Android 13 or later: the remote app ships as an APK in pre-releases. It connects to Pleiad on your PC through your own relay server (docs/remote.md).
See docs/desktop-releases.md for how releases are built, signed, and updated.
Pleiad uses the agents' own CLIs and sign-ins. Install and sign in to the agents you want to use; Pleiad does not issue credentials for them.
See docs/backends.md for per-agent setup.
Run from source
Requires Node.js 22.13 or later.
npm ci
npm start # browser version; open the URL printed on startup
npm run desktop # desktop version
npm test # tests that do not call any LLM
The browser version runs on Windows, macOS, and Linux. The desktop version runs on Windows and macOS.
Command line
bin/pleiad.mjs (the pleiad bin in package.json) talks to a running Pleiad. Its subcommands come from the server's operation list, so new operations appear without changing the CLI.
node bin/pleiad.mjs status
node bin/pleiad.mjs sessions list --limit 5
node bin/pleiad.mjs sessions read <id> --before 2 --after 2
node bin/pleiad.mjs settings get compaction.auto
node bin/pleiad.mjs ops # every available operation
node bin/pleiad.mjs call sessions.get --args '{"sessionId":"<id>"}'
node bin/pleiad.mjs sessions list --json # JSON output
It finds Pleiad through control.json in the data folder (AGENT_HOST_DATA, default ~/.agent-host). Exit codes: 0 success, 2 invalid input, 3 Pleiad is not running, 4 refused or needs the Pleiad screen, 5 other.
Inside a Pleiad conversation the shell already has PLEIAD_CONTROL_URL and PLEIAD_CONTROL_TOKEN, so the CLI acts on that conversation under its approval mode.
pleiad mcp serves the same operations as a stdio MCP server for AI tools outside Pleiad: claude mcp add pleiad -- node <repo>/bin/pleiad.mjs mcp.
Details: Design ("操作の一覧").
Documentation
Detailed documentation is currently in Japanese.
Security
The server listens on 127.0.0.1 by default and requires a token.
To report a vulnerability, see SECURITY.md.
License
Apache License 2.0. Third-party notices are in NOTICE.
The Pleiad name and logo are not covered by this license. If you distribute a fork, use a different name.
