Skip to content

firstintent/ccteam

Repository files navigation

ccteam

A multi-agent orchestrator on top of Claude Code — describe what you want, ccteam picks the command. No YAML. No CLI flags to memorize.

What do you want to do?

Pick the row that matches your goal. The right column is the command — copy-paste into any Claude session.

You want to do                                  → Run this
──────────────────────────────────────────────────────────────────────
Get a feel for a new codebase (60 s, zero deps)  /ccteam-scan --quick
Audit a codebase navigability / large monorepo   /ccteam-scan
Build / fix / refactor (watching it work)        /ccteam-team "<task>"
Cross-vendor second opinion on a hard call       /ccteam-advise "<question>"
A private IM assistant (24/7, always on)         /ccteam-creator "build me a <X> assistant"
A multi-bot IM round-table                       /ccteam-creator "a few bots in a group"
Run a long task overnight (hands-off)            /ccteam-creator "<task>, run while I sleep"
List / pause / resume / check spending           /ccteam-control list | pause | cost
Wire up an IM token (Telegram / Slack / Discord) /ccteam-im-setup
Verify your install / MCP surface / Codex critic ccteam doctor [--verify-mcp | --check-codex-auto-critic | --check-cost-orphan]
Not sure? Just describe it in natural language   /ccteam "<what you want>"

Each command is a Claude Code slash command. Type it in a claude session — /ccteam <NL> is the universal entry; the others let you skip the router when you already know the path.

Get started

# 0. Install Claude Code first: https://code.claude.com/docs/install

# 1. Install the ccteam CLI binary (one line, no Rust toolchain needed):
curl -sSL https://raw.githubusercontent.com/firstintent/ccteam/main/install.sh | sh
#    Installs to ~/.local/bin/ccteam. The script prints a PATH-export
#    hint if that directory isn't on your PATH yet — follow it, then
#    restart your shell (or `source ~/.bashrc` / `source ~/.zshrc`).
#    Pin a specific release tag: CCTEAM_VERSION=<tag> curl ... | sh
#    Install system-wide:        CCTEAM_INSTALL_DIR=/usr/local/bin curl ... | sh
#    Or download an archive from https://github.com/firstintent/ccteam/releases
#    Build from source instead (requires Rust 1.85+):
#      cargo install --git https://github.com/firstintent/ccteam ccteam-cli

# 2. Inside any Claude session, register the marketplace + install the plugin:
claude
/plugin marketplace add https://github.com/firstintent/ccteam
/plugin install ccteam

Inside any Codex session(same two-step works ── ccteam binary must be on $PATH from step 1):

codex plugin marketplace add firstintent/ccteam
# 3. Try the universal entry — describe what you want in any language:
/ccteam "scan this repo and tell me what it does"
/ccteam "fix the TypeScript errors in src/"
/ccteam "build a Telegram bot that summarizes my GitHub PRs at 7am"

# 4. (Optional) Bootstrap a per-project workflow scaffold from the CLI:
ccteam init <project>

Supported platforms for the prebuilt binary: Linux x86_64, macOS arm64 (Apple Silicon), macOS x86_64 (Intel). Windows users: install via WSL2 and use the linux-x64 binary — native Windows isn't supported because tmux + inotify + POSIX signals are foundational to ccteam. On macOS, if Gatekeeper blocks the binary on first run: xattr -d com.apple.quarantine ~/.local/bin/ccteam.

Step 2 registers ccteam as a Claude Code plugin marketplace, then installs the plugin — the seven /ccteam* slash commands and mcp__ccteam__* MCP tools light up immediately. The 5-minute walkthrough for "private IM assistant" (the flagship use case) lives in docs/quickstart.md.

Three ways to talk to ccteam

  • Inside a Claude session/ccteam <NL> is the universal entry; the per-task slash commands above are shortcuts. MCP tools (mcp__ccteam__chat_*, mcp__ccteam__advise_*, mcp__ccteam__admin_*) are the programmatic path for anything a sub-skill does.
  • From IM — DM your bot directly, or @ccteam <NL admin> inside a group for control (pause, cost, list, stop everything, …).
  • Web dashboard (read-only) — http://localhost:7331 to watch workflows, transcripts, and 24h spend.

The AI runs on your computer — it can read your files, run your commands, touch your code. Your phone / IM is just the entry point; close the laptop lid and the workflow keeps running.

Run as a daemon

ccteam start runs the orchestrator in the foreground (or as a detached process). To stop it cleanly:

kill -TERM $(cat ~/.ccteam/ccteam.pid)   # or just Ctrl+C in the foreground terminal

It drains within 5 seconds, releases the web port, and unlinks its pidfile automatically. Long-running tmux chat sessions survive a daemon restart and are re-attached on the next ccteam start — upgrading ccteam does not lose your bot's context. If a chat pane was killed (OOM, manual tmux kill-session), the next ccteam start re-spawns it with claude --resume <name> so the model reloads its full API-level context (tool-use history, cache, reasoning) losslessly via the official Anthropic CLI path. If resume isn't possible (no on-disk session, schema drift), ccteam falls back to a fresh session and emits a visible chat_session_reset event — you'll see the bot acknowledge the reset rather than silently forget.

Docs

What you want Read this
Decision tree — task to command (with examples) docs/task-to-command.md
5-minute walkthrough for the flagship IM-bot use case docs/quickstart.md
Full user manual (every scenario, every flag) docs/user-manual.md
Copy-paste a ready-made use case docs/recipes.md
Something broke docs/troubleshooting.md
Advanced customization / Codex integration docs/advanced/

License & acknowledgements

See LICENSE. Built on Claude Code (the runtime) and openhuman/channels (14+ IM platforms in Rust); orchestration taxonomy from Anthropic's Building Effective Agents; IM-bot pattern (tmux + send-keys + transcript polling) inspired by ccgram and oh-my-claudecode.

About

An unattended multi-agent orchestration engine for Claude Code, empowering an autonomous dev-team to drive software from intent to closed-loop delivery. 基于 Claude Code 的“无人值守”多智能体编排引擎,构建从意图输入到闭环交付的自主开发团队

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors