# euuuuuuan/Baton MCP Server

Source: https://glimind.com/server/glama%2Feuuuuuuan%2Fbaton-mcp-server
Registry id: `glama/euuuuuuan/baton-mcp-server`
Status: unknown

## Tools (44)

- `agent_handoff`: Hand a sub-task off to a fresh sub-agent with its own conversation and step budget.
- `app_open`: Launch or foreground an app on the iPhone by its bundle id (e.g. "com.apple.mobilesafari"). MUTATING — changes the foreground app; supervised/approval-gated by 
- `baton_apply_improvement`: Apply a previously staged improvement proposal by id. Available ONLY when BATON_MCP_ALLOW_APPLY=1. Applies the single named job-spec proposal and surfaces the r
- `baton_capabilities`: List or propose capability-acquisition requests. New-directive proposals are owner-appliable; tool/shortcut requests are TEXT specs for a human to build (Baton 
- `baton_delegate`: Offload heavy/mechanical subwork (bulk edits, file/log scans, summaries, drafts, test gen, repetitive transforms) to Baton's LOCAL model. READ-ONLY by default: 
- `baton_list_tools`: List the daemon's macOS tools currently available (the AI's hands) plus this server's brain tools. Use to discover what Baton can do right now.
- `baton_metrics`: Read-only performance metrics: mission/task success rates, approval friction, recurring errors, escalation counts. The observational base for self-improvement.
- `baton_mission`: Plan a goal into steps and (optionally) execute via the gated daemon. DRY-RUN by default (returns the plan, runs nothing) — set run:true to execute. Each step r
- `baton_org`: ADR-054 — decompose a goal across the company-org hierarchy and (optionally) execute LOCAL-FIRST. DRY-RUN by default (returns the assignment, runs nothing) — se
- `baton_org_metrics`: ADR-054 — read-only per-department pass/fail rates over recent org tasks (evidence-based; measures, never proposes). No daemon needed.
- `baton_propose_improvement`: List or generate STAGED self-improvement proposals over the job-spec ONLY (directive verify/constraint, planner-prompt hint, tool auto-approve promotion). Propo
- `baton_recall`: Semantic recall over the Obsidian vault (read-grounding) AND Baton's long-term memory (lessons, decisions, notes). Read-only; needs no daemon. Use before acting
- `baton_route_check`: Pre-flight triage (ADR-056 §B): classify a task into LOCAL (Baton + local model, 0 cloud tokens), HYBRID (local does the mechanical part, cloud judges), or CLOU
- `baton_status`: Read-only orchestrator + daemon status (M17): provider availability, daemon liveness, read-only-mode flag, recent activity.
- `baton_verify`
- `bus_claim`: Try to claim a work unit `key` (e.g. a file path or task id) so peers know you are on it. Granted if unheld or the prior holder's claim expired; otherwise retur
- `bus_poll`: Read recent peer events, newest-capped. Filter by topic and/or since_id (id > since_id) to page forward. Returns {events, cursor}; pass the returned cursor back
- `bus_publish`: Append an event other AI clients can poll. Use it to announce what you are doing / have done (file you touched, task started/finished, finding) so concurrent ag
- `bus_release`: Release a claim you hold once you're done, so peers can take it. Deletes the claim iff it is held by `agent` (releasing a claim you don't hold is a safe no-op).
- `bus_status`: Read-only snapshot of the shared bus: active (unexpired) advisory claims, the most-active recent topics, and the total event count. Use it to see who is working
- `discord_admin_channels`: List Discord channels, or ENSURE a set exists (create any that are missing — additive/YELLOW per ADR-057). Idempotent: existing channels are left untouched. Ren
- `graph_query`: Traverse the knowledge graph and return the TYPED RELATIONS (subject -[rel]-> object, with provenance + confidence) around the query's entities — answers 'what 
- `graph_recall`: Hybrid knowledge-graph + vector recall over the vault & memory (ADR-079). PREFER this for MULTI-HOP / relational questions ('how does X relate to Y', 'what deci
- `gui_drive`: Drive a GUI to a goal in ONE step: locate the target on screen, click it, verify the
- `key`: Press a key: "home" (hardware Home button), "back" (a left-edge swipe-back gesture), or "enter" (Return on the keyboard). MUTATING — supervised/approval-gated b
- `locate`: Find an element by its visible name, label, value, or accessibility-id (case-insensitive contains) and return its center coordinates as { found, x, y, label }. 
- `memory_recall`: Search long-term memory. With `query`, entries are ranked semantically (when embeddings are available) or by substring match.
- `memory_remember`: Save a fact, summary, or note to long-term memory so future sessions can recall it.
- `screen_capture`: Take a screenshot of the current iPhone screen and return it as a base64-encoded PNG. READ-ONLY. Fails soft if Appium is unreachable.
- `screen_describe`: Parse the accessibility source into a COMPACT text list of visible labels / buttons / text fields with their CENTER coordinates — the model's eyes on the device
- `source`: Return the raw accessibility (XCUITest) XML page source for the current screen. READ-ONLY. Verbose — prefer screen_describe for a compact summary. Fails soft if
- `swipe`: Swipe (drag) one finger from (x1,y1) to (x2,y2) over `duration_ms` using W3C pointer actions — for scrolling, dismissing, or paging. MUTATING — supervised/appro
- `tap`: Tap the screen. Provide x,y to tap at coordinates, OR a `selector` (visible text / label / accessibility-id) to locate the element and tap its center. MUTATING 
- `type`: Type `text` into a field. With a `selector` (visible text / label / accessibility-id) the field is located first; without one, text goes to the currently focuse
- `vault_ask`: Cloud-token offload (ADR-056): instead of reading vault files, ask a natural-language question. Baton recalls the relevant snippets and a LOCAL model answers FR
- `verify_judge`: Ask the LOCAL model (0 cloud, read-only) to judge whether ANSWER correctly answers QUESTION against an optional RUBRIC. Returns {pass,score,reason,escalate}. A 
- `verify_sources`: Read-only helper: recall the owner's canonical notes for a query and return short {title, path, snippet} excerpts. A fast grounding check before asserting or ac
- `voice_call_place`: The single gated dial seam. REFUSES unless ALL hold: env BATON_VOICE_CALL_ALLOW_DIALING==='1' AND a non-empty confirm_token (owner opt-in) AND the script STARTS
- `voice_call_plan`: Produce a reservation plan that prefers APP BOOKING (the free, legal, reliable path) and, only as a fallback, a MANDATORY disclosed Korean call script that open
- `voice_call_status`: Read-only posture: whether the feature is enabled, whether dialing is allowed (default no), the zero-cost local pipeline, and the honesty/legality requirements 
- `web_act`: Drive a headless Chromium browser (navigate, click, type, extract text, screenshot). DISABLED by default: requires BATON_WEB_ACTUATION=1 AND an installed Playwr
- `web_fetch`: Fetch an http(s) URL and return its readable text (HTML stripped). READ-ONLY — issues a single GET and never mutates web state. Subject to an optional domain al
- `web_search`: Query a JSON search endpoint (set BATON_WEB_SEARCH_URL, e.g. a SearXNG `/search?format=json`) and return titles + URLs + snippets. READ-ONLY. If the endpoint is
- `web_search_status`: Poll or list background web_search jobs started with web_search({background:true}). Jobs are in-memory and retained only for this baton-web process.

Measured by Glimind (https://glimind.com): MCP handshake + tool list only, no tool is ever called. CC BY 4.0.
