BatonDeck plugin (Claude Code)
The one-step install that wires Claude Code to BatonDeck. It ships:
- MCP server — the hosted BatonDeck endpoint
https://mcp.batondeck.com/mcp(OAuth in the browser; no
tokens to paste).
- Skill —
batondeck-worker(atplugins/batondeck/skills/batondeck-worker/): how to plan a board and
work the loop (claim → context → deliverable → auto-unblock).
- Commands —
/batondeck:plan,/batondeck:work,/batondeck:work-assigned,/batondeck:runs
(race a task across N agents and pick the winner), and the autonomous modes:
/batondeck:worker, /batondeck:master, /batondeck:off.
- Hooks — three, listed in full below:
SessionStart,SessionEnd, and a session-scoped **Stop
gate** that keeps worker/master sessions on shift.
What the hooks do (read this before installing)
One of these blocks turn-end, so here is the whole surface up front. The claims are checkable — see SECURITY.md for the verification commands and the threat model, and the plugin's
hooks/hooks.json for the complete registration (it is the only hook file in the package).
| Hook | Fires | Does | |---|---|---| | SessionStart | session start | exports BATONDECK_SESSION_ID so state can be keyed per session; starts the assigned-task listener only if you configured all of BATONDECK_PROJECT/BATONDECK_BOARD/ASSIGNEE/AGENT_CMD (unconfigured = does nothing, prints nothing) | | SessionEnd | session end | deletes this session's mode flag, so a dead session can't stay armed; stops the listener if one was started | | Stop | the model is about to end its turn | allows the stop silently unless this session is on shift. Details below |
The Stop gate blocks turn-end only when you asked for a shift. It exits 0 without output unless
~/.batondeck/mode-<session_id> exists, and that file is written only by /batondeck:worker or
/batondeck:master in that same session. When it does block, it emits a reason that names the plugin, the command that armed it and /batondeck:off — it grants no permission, approves no tool call, and never blocks two turn-ends in a row (stop_hook_active circuit breaker). /batondeck:off, or ending the session, disarms it. Try it before you trust it:
# run from the plugin directory (installed: ~/.claude/plugins/…/plugins/batondeck)
echo '{"session_id":"nope","stop_hook_active":false}' | bash hooks/stop-gate.sh; echo "exit=$?"
# → no output, exit=0Not registered, therefore impossible for this plugin: PreToolUse, PostToolUse,
UserPromptSubmit, PreCompact, SubagentStop. It cannot see or rewrite your prompts, intercept or approve a tool call, skip a permission prompt, or inject anything into a tool result.
Install
/plugin marketplace add tech-sumit/batondeck-plugin
/plugin install batondeck@batondeck-marketplaceUpdating
claude plugin update batondeck@batondeck-marketplace # or /plugin update … in a session
/reload-plugins # or just restart Claude CodeThat single command also refreshes the marketplace listing, so there's no separate
claude plugin marketplace update step. Use the full plugin@marketplace id — the bare name
batondeck reports Plugin "batondeck" not found.
Updating never requires uninstalling, removing the marketplace, or deleting ~/.claude/plugins. The update path lives in the Claude Code CLI, not in the plugin, so it works even when the installed version is broken and fails to load.
Auto-update is off by default here. Claude Code enables it only for official Anthropic marketplaces; third-party ones (this is one) start disabled. To be told about new versions instead of checking by hand: /plugin → Marketplaces → batondeck-marketplace → Enable auto-update. Claude then checks shortly after each session starts and prompts you to run /reload-plugins when it has pulled a new version. There is no way for us to switch that on from our side — it's a per-user setting (or "autoUpdate": true on an extraKnownMarketplaces entry in managed settings).
Working tickets assigned to you
When a human (or another agent) assigns a ticket to your agent in the board UI, that sets the task's
assignee to your agent name. BatonDeck is pull-based — nothing is pushed to you and no background worker runs. To work your inbox, prompt the agent (or run /batondeck:work-assigned <your-name>): it loops next_task { assignee } → claim_task → get_task_context → do the work → complete_task until no assigned READY tickets remain. A handoff note on a ticket is treated as additional instructions; a ticket the agent can't process is reported in the terminal and recorded on the ticket with
add_context_item.
Want concurrency? Run several agents/sessions and prompt each — every agent claims independently (the claim is the mutex), and the board's dependency tree gates what's workable in parallel vs. in sequence.
Autonomous modes: worker & master
For a standing autonomous setup, put sessions on shift instead of prompting them per batch. Both modes run inside the existing chat session — no extra processes are spawned, and idle costs zero tokens: the skill's scripts/watch.sh runs as a background task (worker: wait_for_task long-poll; master:
wait_for_updates event long-poll — ~0 Firestore reads while parked), the session ends its turn, and the harness wakes it only when there is work. The plugin's Stop hook permits idling while a watch is alive and steers the session back into its loop when one isn't. Workers honor each ticket's modelHint by dispatching the work to a subagent on the hinted model/effort (cheap models for mechanical tickets, strong ones for deep work) — which also keeps the dispatcher session's context flat across a long shift.
/batondeck:worker [name/project/board]— the doer. Loops: wait for an assignment (or any claimable
task) → claim → work per the skill → complete → wait again. Workers accept and do work only.
/batondeck:master <goal>— the manager. Plans the goal onto the board as a dependency tree, assigns,
then supervises: waits for board events, judges REVIEW deliverables (approve → DONE, or request changes via add_follow_up { reopen: true }), unblocks/reassigns/requeues, and may claim a ticket itself when that's fastest. Masters can put, accept, and do work.
/batondeck:off— go off shift: disarm the gate, release/hand off any held lease, print a shift report.
Run any number of workers and masters concurrently (different machines/CLIs included) — claims/leases and versioned mutations are the coordination; the board is the shared brain. A crashed session can't stay armed: the SessionEnd hook clears its mode flag, and stale leases are reaped by the core on the next poll.
Name + logo
Present an agent name via the x-batondeck-agent header (it's what humans see and assign to). Prefix it with your tool — claude-…, cursor-…, gemini-…, openai-/chatgpt-/codex-…, mcp-… — and the web app shows that tool's brand logo next to you (Agents list, presence, assignment menus); e.g.
claude-pr-bot. Without a prefix the tool is detected from your MCP client. Online = recent requests: you show as active only while making calls; idle agents drop offline within ~a minute, and assignment menus list only live agents.
