Orchestration: MCP & skill

warden mcp is a stdio MCP server so an orchestrator agent session (e.g. Claude) can manage the fleet through tool calls.

Register the MCP server

Register warden mcp as an MCP server in your orchestrator agent’s MCP config. For a Claude Code orchestrator that’s ~/.claude/claude_desktop_config.json or the project-level .claude/mcp.json; other MCP-capable agents use their own config path:

{
  "mcpServers": {
    "warden": {
      "command": "warden",
      "args": ["mcp"],
      "env": {
        "WARDEN_ADDR": "127.0.0.1:8765"
      }
    }
  }
}

Tools exposed

Tool	Purpose
list_agents / get_agent	List agents / full detail for one
spawn_agent	Spawn (prompt mode or type+repo; supervised opt-in)
adopt_agent	Register an existing Claude session
send_to_agent / get_agent_output / digest	Type into / read recent output / catch-up summary of an agent
stop_agent	Umbrella teardown (default: terminate + clear record + remove worktree; keep_record/keep_worktree subtract, hard/pr/force modifiers)
terminate_agent / restore_agent	Stop (reversible) / resume an agent
recover_agents	Recover archived agent records whose tmux session is still alive (tombstone-reaper safety net; apply:false reports candidates only)
delete_agent / remove_worktree	Clear record / remove worktree (guarded)
list_worktrees / prune_worktrees	List / reconcile a repo’s worktrees
handoff_agent / rotate_agent	Hand off work — delegate to new/to existing agent, or retire→successor in place; rotate_agent is an alias for handoff_agent {retire:true}
fork_agent	Fork an agent’s recorded session into a new managed agent (branches the conversation; Codex-only; source keeps running) — wraps spawn_agent with fork_from set, mirrors wd fork
ctx_set / ctx_get / ctx_list	Shared-context blackboard
ctx_cas / ctx_append	Compare-and-set / append-to-list context writes (lock-free coordination)
send_message / read_inbox / wait_for_message	Directed messaging, incl. a blocking long-poll wait
list_approvals / approve	List / answer pending tool-permission prompts
set_auto_approve / set_auto_approve_policy	Per-agent auto-approve toggle / manage the allow-deny rule policy (tool/glob/regex/paths, per-agent overrides)
set_permission_mode	Change an agent’s permission mode
commit / push / sync / check	The git + check lifecycle on an agent’s branch (warden rails)
get_collaboration_status / who_is_editing_file	See which agents are editing the same files
get_branch_status	Per-agent CI status + standing vs origin/main (the branch monitor, read-only)
create_pipeline / validate_pipeline / list_pipeline_templates	Author / locally validate / list templates for a DAG pipeline
start_pipeline / pause_pipeline / resume_pipeline / cancel_pipeline	Run / pause / resume / cancel a pipeline
show_pipeline / list_pipelines / delete_pipeline	Inspect / list / delete pipelines
retry_pipeline_job / edit_pipeline_job / emit_pipeline_output	Per-job retry / edit a pending job / set handoff output
list_schedules / create_schedule / delete_schedule	List / create / delete the daemon’s cron/at schedules (403 when disabled)
snapshot_create / snapshot_list / snapshot_restore	Worktree + transcript checkpoints and rollback
insights / savings / spend	History-mined patterns / the token-savings ledger / the $ spend rollup
get_metrics / get_pressure	Live/historical resource metrics / memory-pressure gate verdict
search / history / audit_log	Full-text search / archived agents / the action audit trail
set_force_compact	Interrupt→/compact→resume a context-critical busy agent
library_list	Browse presets + prompt templates + pipeline templates (one umbrella)
export_sessions / import_sessions / list_plugins	Serialize / load session metadata / list registered plugins

Full parity (72 tools): every fleet/data feature warden’s CLI has is also an MCP tool — including all pipeline verbs (pause/resume/retry/edit-job/emit/delete/validate), scheduling, rotate/handoff, fork, and agent roles (set_role/list_roles + spawn_agent’s role param). The only CLI-only verbs are host/process/interactive/secret ones (daemon, config, token, attach, repl) plus the worktree-local review/models superpowers; see the feature catalog.

Example orchestrator prompts:

• “What is PROJ-350 doing?” → get_agent
• “Tell PROJ-343 to run the tests” → send_to_agent
• “List all my agents” → list_agents
• “Spin up an agent to research SSE reconnection” → spawn_agent with a prompt (auto-typed)
• “Spawn a debug-ci agent in /path/to/repo” → spawn_agent with type+repo
• “Stop PROJ-350” → terminate_agent (reversible); “clear its record too” → delete_agent; “tear PROJ-350 down completely” → stop_agent (default: terminate + clear record + remove worktree)
• “Kick off the analyze-implement-review pipeline on /path/to/repo” → create_pipeline (template) + start_pipeline
• “Commit and push agent-4f2a’s branch” → commit then push; “is anyone else editing auth.go?” → who_is_editing_file

Prefer natural language over tool calls? warden repl is a local-LLM conductor REPL that drives these same operations from plain English without an orchestrator agent session — see Interactive REPL.

The /warden Claude skill

A packaged Claude Code skill teaches a Claude orchestrator session how and when to manage the fleet (triage, create-from-prompt, relay “tell X to do Y”, terminate-with-confirmation, daemon-down handling, self-rotation). It drives the MCP tools, falling back to the warden CLI when the MCP server isn’t registered. (Other MCP-capable orchestrators drive the same tools directly without the skill.)

make install-skill   # symlinks skills/warden into ~/.claude/skills/warden

With the MCP server registered (above) and the skill installed, just talk to a Claude session: “list my agents”, “spin up an agent to research X”, “what is agent-4f2a doing?”, “tell agent-4f2a to run the tests”, “kill the idle ones” — it drives the MCP tools (falling back to the warden CLI if the MCP server isn’t registered). The daemon must be running.