# Muster: The Agentic Fire Drill

Canonical product: https://muster-fire-drill.vercel.app/

Source: https://github.com/Arnie016/muster-webmcp

Devpost: https://devpost.com/software/muster-the-agentic-fire-drill

License: MIT

## What Muster is

Muster is a browser-based tabletop fire-drill command room for a building Fire Safety Manager and an exercise team. A human facilitator and an agent work on the same visible training plan. They can read the fictional building file, introduce scripted conditions, inspect occupants and roles, compare the exercise's route options, record facilitator-confirmed actions, find responsibilities without owners, and prepare an after-action draft.

The product is deliberately not a generic chat assistant. Tool calls read or update one deterministic in-page state and leave visible receipts. The human can inspect the request, delegation path, structured input, structured result, purpose, screen change, duration, and safety boundary. Human report approval is a separate control and is not exposed as a tool.

## Intended audience and task

The primary user is a Fire Safety Manager or exercise facilitator preparing a controlled tabletop rehearsal. The primary task is to keep the plan, fictional occupant register, roles, scripted conditions, team decisions, route evidence, and after-action review in one inspectable sequence.

The current submission demonstrates one complete Floor 07 office rehearsal. Floor 03 and Floor 12 are labelled reference plans, not executable scenarios.

## Thirty-second judge path

1. Open https://muster-fire-drill.vercel.app/.
2. Select **Start scenario**, then use the visible **Next action** control to advance one declared tool call at a time.
3. Watch the authored detector signal appear and Stair B become unavailable in the fictional scenario.
4. Select **3D rooms · F07**. Inspect Studio, focus an equipment marker, then step through the room-to-stair walkthrough. A blocked Stair B preview stops before the landing.
5. Ask a concrete question in the Commander conversation, such as where the manual call point is. Inspect the actual tool receipt attached to the turn.
6. Open a Live trace row to inspect the input, result, stated purpose, and visible change. Continue the guided scenario to prepare the review draft; only the human can approve it.
7. Select **Print pack** for the two-page A3 plan and review sheet, vector SVG, or 5500 × 4250 PNG.

The displayed distances are fictional plan fixtures. A drawn route is not an approved evacuation route.

## WebMCP model

Muster registers one manager and 19 bounded page tools through `document.modelContext.registerTool` when the API is available. `run_drill_manager` accepts a high-level exercise intent and delegates to plan, people, equipment, or review specialists. The remaining tools expose narrowly scoped operations over the same visible page.

Current native proof uses Chrome 152 with WebMCP testing enabled. `document.modelContext.getTools()` discovered all 20 tools. `document.modelContext.executeTool()` then ran the declared fictional rehearsal: plan read, manager orientation, scenario start, zone inspection, exit change, route comparison, three facilitator-confirmed records, coverage check, and manager-prepared review. The same page reached “Ready for human review”; the approval control remained untouched. This proves discovery and a deterministic fixture sequence in that tested production tab. It does not prove autonomous planning, support in every browser, or every ChatGPT surface.

## Tool groups

### Manager

- `run_drill_manager`: route a high-level request to a bounded specialist and named page tool.

### Plan

- `read_plan`: read the selected building, floor, exits, roles, revision, and visible fixture counts.
- `start_drill`: start the fictional tabletop scenario.
- `inspect_zone`: focus one plan zone and return its fictional people, assistance need, and nearest-exit context.
- `compare_routes`: compare scripted distance and availability; optional `view: "3d"`, `preview_exit`, and `checkpoint` show a bounded five-checkpoint walkthrough. Previewing is not recording a decision or certifying an evacuation route.
- `analyze_route_sketch`: measure a facilitator-drawn path and check its endpoint against the scenario's exit state.
- `read_site_context`: read fictional setting, assembly areas, service access, and mapping boundaries.
- `read_room_profile`: inspect one room's use, plan-derived dimensions, located equipment, and exercise context; `view: "3d"` focuses the room in the Floor 07 cutaway.

### People and scenario

- `prepare_team_handoff`: preview a known fictional responder, destination room and task on the shared 3D plan. Returns a diagram path, rationale checks and `requires_human_confirmation`. Human-only confirmation updates the assignment record, 2D/3D markers, roster and native readback. No arrival, clearance, assistance completion or dispatch is inferred. Older previews become invalid after scenario changes.

- `send_inject`: add an authored complication to the exercise.
- `record_action`: preserve a facilitator-confirmed team action and owner.
- `read_floor_register`: read aggregate fictional zone counts and assistance ownership.
- `read_status_board`: show which exercise records exist without inferring floor clearance.
- `record_human_signal`: record an observed confirmation, uncertainty, disagreement, or delay without inferring intent or competence.

### Equipment and guidance

- `read_drill_guide`: read the before, during, or after tabletop checklist.
- `read_hazard`: read the authored scenario phase without claiming a physical fire-spread model.
- `read_equipment`: read planned fixture and inspection records without certifying adequacy; optional `item_id` and `view: "3d"` focus a named marker. A manual call point is inventory, not a button that raises an alarm.
- `read_lessons`: read dated fictional exercise findings and the changes they motivated.

### Review

- `check_coverage`: find active exercise problems without a recorded owner.
- `stage_report`: prepare a reviewable after-action draft. It cannot approve the draft.

## Product evidence

- The 18-floor building is an orbitable WebGL training twin with direct floor selection and a CSS fallback.
- Floor 07 is a pan-and-zoom response plan with dimensions, rooms, route state, equipment, fictional occupant groups, responder markers, and inspectable profiles.
- Floor 07 also has a room-level cutaway with original procedural desks, seating, equipment, doors, inspection markers, perspective orbit, orthographic top view, and a checkpoint-follow camera. Geometry is derived from the fictional plan; height, furniture, and exact marker positions are illustrative, not surveyed.
- Native WebMCP tests additionally focus Studio, inspect the manual call point, and request a blocked Stair B checkpoint. The visible preview clamps before the unavailable landing and does not record an action.
- The print pack snapshots the same scenario and route into two A3 pages. Accounted-for counts and reviewer signoff stay blank unless actually recorded. SVG and 5500 × 4250 PNG export are available; high resolution does not establish engineering accuracy.
- An animated orange authored-signal trail shows the scripted scenario changing. It is explicitly not a sensor event or physical spread prediction.
- The guided flow advances one declared change at a time and explains what changed and what the facilitator should inspect next.
- Browser checks cover desktop, 390 by 844 mobile, the WebGL twin, route analysis, the complete rehearsal, and human report approval.
- Deterministic tests include six people-data tests and 500 shuffled workflows for repeat-safe actions, invalid transitions, route boundaries, and approval gates.

## WebMCP Challenge fit

- **WebMCP leverage:** the agent discovers structured tools and operates the same live page as the facilitator. Calls produce visible plan, people, route, scenario, trace, and review changes rather than disconnected chat text.
- **Execution:** the public product supports a complete guided rehearsal on desktop and mobile, an interactive WebGL/CSS spatial view, inspectable tool contracts and receipts, deterministic tests, and native discovery and execution evidence.
- **Potential impact:** Muster targets the real coordination shape of tabletop building drills: changing conditions, named responsibilities, route evidence, and after-action review. No field adoption, safety outcome, or professional validation is claimed yet.
- **Creativity and ambition:** one manager coordinates four bounded specialist desks and 19 page tools over a spatial command room while preserving an explicit human approval boundary.

## Safety and truth boundaries

Muster does not monitor a real building, read sensors, infer live occupancy, certify equipment, calculate physical fire spread, approve an evacuation route, raise alarms, control doors, contact emergency services, dispatch responders, or provide instructions during a live emergency.

Every person, role, address, building, count, room, dimension, distance, equipment record, inspection, timestamp, and incident in the current product is a fictional training fixture. A displayed ownership ratio is exercise-action coverage, not a safety or readiness score.

Suggested agent behavior:

- Read the selected plan before changing the scenario.
- State which fictional fixture or facilitator observation supports each conclusion.
- Use route tools only for the scripted exercise and preserve the qualified-human review boundary.
- Record human signals only when the facilitator reports what was observed.
- Stage the report only after checking active-problem ownership.
- Never claim that silence proves agreement, a record proves clearance, or a plan symbol proves equipment adequacy.
- Stop using Muster during a real emergency and follow the official local emergency process.

## Authoritative public pages

- Product: https://muster-fire-drill.vercel.app/
- Markdown product summary: https://muster-fire-drill.vercel.app/index.md
- Short agent index: https://muster-fire-drill.vercel.app/llms.txt
- Operator skill: https://muster-fire-drill.vercel.app/SKILL.md
- Operator guide: https://muster-fire-drill.vercel.app/docs/OPERATOR_GUIDE.md
- Source and tests: https://github.com/Arnie016/muster-webmcp
- Native receipt: https://github.com/Arnie016/muster-webmcp/blob/main/docs/screenshots/muster-native-webmcp.png
- Public support: https://github.com/Arnie016/muster-webmcp/issues