subagent-cockpit-mcp
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| SUBAGENT_COCKPIT_HOME | No | Moves the whole ~/.subagent-cockpit directory (where config.json and session data live). | |
| SUBAGENT_COCKPIT_WEBUI_PORT | No | The port for the read-only Web UI (default 7200). Overrides the webui.port value in the config. 0 turns the web UI off. |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
This server publishes no instructions, or was last inspected before Glama recorded them.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| subagentsB | List configured subagents. |
| sessionsA | List sessions on disk, most recent first, with state, title, last_prompt and last_at. Use it to find a session id after an interrupted open. |
| openA | Open a session: reconnect when |
| promptA | Send a task to an IDLE session. In ASKING or QUESTIONING it returns the pending approval or question instead. A session closed for idleness is reopened first. Blocks until the turn is done, needs an approval or asks a question; the server sets no time limit. If your client times out, the turn keeps running and the connection stays: call |
| approveA | Allow this one tool call (allow_once), or pick option_id explicitly. Blocks until the turn is done, needs an approval or asks a question; the server sets no time limit. If your client times out, the turn keeps running and the connection stays: call |
| allowA | Allow and stop asking for similar calls in this session only. Uses the agent's session-scoped option when it offers one; otherwise this server allows once and answers matching requests itself until the session closes. Never writes a persistent agent rule. Blocks until the turn is done, needs an approval or asks a question; the server sets no time limit. If your client times out, the turn keeps running and the connection stays: call |
| rejectA | Reject this tool call (reject_once). Codex and Claude Code continue the turn; Grok and OpenCode end it as cancelled. Blocks until the turn is done, needs an approval or asks a question; the server sets no time limit. If your client times out, the turn keeps running and the connection stays: call |
| answerA | Answer a question the agent asked (question_needed). requested_schema is the agent's own form, passed on as sent: with action "accept" (the default), give content as an object whose keys and values follow it. "decline" refuses to answer, "cancel" dismisses the question. Blocks until the turn is done, needs an approval or asks a question; the server sets no time limit. If your client times out, the turn keeps running and the connection stays: call |
| cancelA | Cancel the current turn. Returns "cancelling" until the agent confirms, or "done" when already idle. |
| set_configA | Change a session config option (model, reasoning_effort, mode, ...). Returns the full config_options. Unknown ids or values fail with the available choices. A session closed for idleness is reopened first. |
| statusA | Session state, cancelling flag, current config values, pending approvals and questions, and whether a blocking call is waiting. Cheap to call repeatedly; the config choices are in open, set_config and snapshot. A session closed for idleness reports CLOSED with its last values. |
| progressA | The agent's latest execution feedback: its plan, context usage, session title and slash commands. Not needed to run a task; call it only when you want to know. Cheap, never blocks. |
| outputC | Final reply of the most recent turn. |
| closeA | Close a session, or all sessions of this server when session is omitted. Session history is kept. |
| deleteC | Delete a session and its native agent session (deleted / archived / unsupported). |
| snapshotA | Full conversation of open sessions, the same as the webui shows: prompts, replies, thinking, tool calls with their input, output and diffs, plans, approvals, questions and agent-specific messages, plus each session's state and pending notifications. It can be large: pass session, and use last, types or exclude_types to keep it small. Pass back server_id and seq as since to get only what changed. ack is for UI clients that relay notifications. UI clients that poll: pass simplify to leave out tool call details (input, output, content) and session commands, and get one item in full with item. |
| itemA | One conversation item in full, by the id snapshot gave it: a tool call with its input, output and diffs, or any other item. For UI clients that poll snapshot with simplify and show details on demand. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 17 tools
Most tools target clearly distinct lifecycle or interaction steps: open/close/delete sessions, prompt, approvals, questions, status, output, and snapshot. Two minor overlaps remain: approve vs allow both permit calls (differing only in one-time vs session scope), and item is essentially a detail view of snapshot. The descriptions do enough to separate them, but an agent could occasionally hesitate.
Names are consistently lowercase snake_case, with only set_config using a multiword form and the rest single tokens. The set mixes action verbs (open, prompt, approve, cancel) with resource/state nouns (subagents, sessions, status, output, snapshot, item), but the style is uniform and readable. It is not a strict verb_noun convention, though it is internally consistent.
Seventeen tools is slightly heavy but reasonable for a session cockpit covering lifecycle, prompts, approvals, questions, config, and conversation inspection. A few tools (progress, item) are optional or UI-polling helpers and could be consolidated, but each has a plausible role. This sits at the low end of the heavy range rather than being overstuffed.
The surface covers the main session lifecycle: list subagents/sessions, open, prompt, approve/allow/reject, answer questions, configure, cancel, inspect status/output/progress, snapshot conversation, and delete. Gaps are minor, such as no tool for defining or editing subagents and no session search/rename operation. Agents can work around these for the core cockpit purpose.