agentxchain

CLI for multi-agent coordination in your IDE. Define a team of AI agents, let them take turns building your project — each in its own IDE window.

Install

npm install -g agentxchain

Or run without installing:

npx agentxchain init

Quick start

Happy path: net-new project

agentxchain init
cd my-project
agentxchain kickoff

Happy path: existing project

Run these commands from inside your existing project folder:

agentxchain doctor
agentxchain generate
agentxchain kickoff

Each agent runs in its own Cursor window for a single turn at a time. The referee loop (watch / supervise --autonudge) determines the next agent and wakes that specific session.

Agents are now required to maintain TALK.md as the human-readable handoff log each turn.

Commands

Full command list

agentxchain init
agentxchain status
agentxchain start
agentxchain kickoff
agentxchain stop
agentxchain config
agentxchain rebind
agentxchain generate
agentxchain watch
agentxchain supervise
agentxchain claim
agentxchain release
agentxchain update
agentxchain doctor
agentxchain validate

IDE options

agentxchain start                   # Cursor (default) — one window per agent
agentxchain start --ide vscode      # VS Code — uses .agent.md custom agents + hooks
agentxchain start --ide claude-code # Claude Code — spawns CLI processes

Additional flags

agentxchain kickoff                # guided first-run PM-first workflow
agentxchain kickoff --ide vscode   # guided flow for VS Code mode
agentxchain kickoff --send         # with Cursor auto-nudge auto-send enabled
agentxchain kickoff --interval 2   # nudge poll interval override
agentxchain kickoff --no-autonudge # skip auto-nudge prompt

agentxchain start --agent pm        # launch only one specific agent
agentxchain start --remaining       # launch all agents except PM (PM-first flow)
agentxchain start --dry-run         # preview agents without launching
agentxchain validate --mode kickoff # required before --remaining
agentxchain validate --mode turn --agent pm
agentxchain validate --json         # machine-readable validation output
agentxchain watch --daemon          # run watch in background
agentxchain supervise --autonudge   # run watch + AppleScript nudge loop
agentxchain supervise --autonudge --send   # auto-press Enter after paste
agentxchain supervise --interval 2  # set auto-nudge poll interval
agentxchain rebind                  # regenerate agent prompt/workspace bindings
agentxchain rebind --open           # regenerate and reopen all Cursor agent windows
agentxchain rebind --agent pm       # regenerate one agent binding only
agentxchain claim --agent pm        # guarded claim as agent turn owner
agentxchain release --agent pm      # guarded release as agent turn owner
agentxchain release --force         # force-release non-human holder lock

macOS auto-nudge (AppleScript)

If you want the next agent chat to be nudged automatically when turn changes, use the built-in AppleScript helper.

1. Keep watcher running in your project:

agentxchain watch
# or use the combined command:
agentxchain supervise --autonudge

2. In another terminal (from cli/), start auto-nudge:

bash scripts/run-autonudge.sh --project "/absolute/path/to/your-project"

By default this is paste-only (safe mode): it opens chat and pastes the nudge message, but does not press Enter.

3. Enable auto-send once confirmed:

bash scripts/run-autonudge.sh --project "/absolute/path/to/your-project" --send

Stop it anytime:

bash scripts/stop-autonudge.sh

Notes:

• Requires macOS (osascript) and jq (brew install jq)
• Grant Accessibility permissions to Terminal and Cursor
• The script watches .agentxchain-trigger.json, which is written by agentxchain watch
• run-autonudge.sh now requires watch to be running first
• The script only nudges when it finds a unique matching agent window (no random fallback)

How it works

Cursor mode (default)

1. agentxchain kickoff launches PM first for human-product alignment
2. Each window gets a unique prompt copied to clipboard
3. Kickoff validates PM signoff and launches remaining agents
4. Agent prompts are single-turn: claim → work → validate → release → stop
5. Agents use the latest Next owner: in TALK.md to pick who goes next (fallback: config order)
6. Human can claim to pause and release to resume anytime

VS Code mode

1. agentxchain init generates .github/agents/*.agent.md (VS Code custom agents) and .github/hooks/ (lifecycle hooks)
2. VS Code auto-discovers agents in the Chat dropdown
3. The Stop hook acts as referee — hands off to next agent automatically

Turn ownership

Agent turns are handoff-driven:

• Each turn appends a Next owner: in TALK.md with a valid agent id
• watch/supervise dispatches the next trigger from that handoff
• claim --agent <id> enforces that expected owner (with guarded fallback)

Key features

• One window per agent — each agent has its own Cursor window and chat session
• Referee-driven coordination — watch/supervise wakes the next correct agent each turn
• Works in Cursor, VS Code, Claude Code — adapters for each IDE
• User-defined teams — any number of agents, any roles
• No API keys or cloud required — everything runs locally
• Human-in-the-loop — claim/release to intervene anytime
• Team templates — SaaS MVP, Landing Page, Bug Squad, API Builder, Refactor Team
• Lock TTL — watch can force-release stale locks as a safety net

VS Code extension (optional)

For a richer UI in VS Code:

code --install-extension cli/vscode-extension/agentxchain-0.1.0.vsix

Adds: status bar (lock holder, turn, phase), sidebar dashboard, command palette integration.

Publish updates (maintainers)

cd cli
bash scripts/publish-npm.sh              # patch bump + publish
bash scripts/publish-npm.sh minor        # minor bump + publish

If NPM_TOKEN exists in agentXchain.dev/.env (project root), the script uses it automatically.

Links

• agentxchain.dev
• GitHub
• Protocol v3 spec

License

MIT