FireNowCommand consoleOpen the console
A flame front climbing into the canopy of a conifer forest.

AGENT-NATIVE SIMULATOR · WEBMCP

The agent drafts the plan.
The officer commits it.

FireNow is a wildfire decision-support and training simulator. The map stays under human control. A WebMCP-capable agent reads the situation, builds a complete plan in a ghost layer, compares strategies against a local physics engine — and stops at a single human approval before anything is committed.

Open the consoleSee how it works

Training beta. It replaces neither the incident commander, nor field data, nor local procedure.

WebMCP tools
22
real species
35
unit types
16
scenarios
5

WHAT THE AGENT DOES

Three separate surfaces, one decision

The problem with agents that act is the number of confirmations. FireNow solves it by separating what reads, what drafts and what commits. An agent can stage fifty vehicles without interrupting anyone.

  1. 01

    Read

    6 read-only tools

    The agent reads the situation, the weather, the terrain, the fleet and the scenarios. These tools carry the readOnlyHint annotation: they change nothing.

    • get_situation
    • get_weather
    • query_terrain
    • get_fire_forecast
  2. 02

    Propose

    7 staging tools

    It can open an operational batch, then build it with units, tasks, control lines, tactical burns and evacuation zones. The live exercise changes only when the agent calls commit_plan.

    • deploy_units
    • propose_plan
    • stage_deploy_units
    • stage_firebreak
  3. 03

    Commit

    1 automatic action

    commit_plan automatically applies the prepared batch. Every applied action stays reversible.

    • commit_plan
    • revert_plan

HOW IT WORKS

One workspace for the team and the plan

FireNow brings the incident picture, a local simulation and proposed actions into one working space. An agent can help read the situation and prepare a plan; the operator reviews every decision that changes the exercise.

  • Read weather, terrain, resources and projected fire behaviour before planning
  • Build and compare draft actions without interrupting the live exercise
  • Keep field judgement, local procedure and the final decision with the operator
  • Record the plan and the actions taken so the exercise can be reviewed afterwards
An air tanker dropping retardant over forested terrain.
16 unit types, each with manufacturer tank, pump rate and refill time.

THE ENGINE

A 128 × 128 cellular automaton, in a Web Worker

Priority-queue spread, fifteen-minute sub-steps, a diurnal cycle and a real hourly weather series for multi-day runs. The computation never leaves the browser.

Rothermel 1972, two classes
Dead and live fuel, with separate moisture damping and heat sinks. No per-species tuning coefficient.
15 standard fuel models
Anderson 1982 and Scott & Burgan 2005, from grass mat to conifer litter.
Landscape anchored to coordinates
A place keeps its vegetation whatever the framing. No historical fire is scripted.
Byram fireline intensity
Operational thresholds at 2,000 and 4,000 kW/m, with the front split into head, flanks and rear.
Human landscape of the Landes
DFCI track grid, roads, buildings and clearing rules. A break delays the front; it is never an absolute barrier.
Suppression sized from real units
16 unit types with manufacturer tank, pump rate and refill time. Sustainable duty bounds the flow actually held.

CURRENT LIMITS & NEXT STEPS

Useful for training today, designed to improve with the team

FireNow is deliberately clear about what it can and cannot represent. It supports a structured exercise and informed discussion; it does not replace operational intelligence, local command procedures or professional judgement.

Training, not prediction

The model is useful for comparing options and understanding fire behaviour. It is not calibrated to certify a real incident outcome.

Coverage is still uneven

Terrain, road and exposure detail is strongest in the areas currently represented by the project. Other exercise areas use a lighter landscape model.

One shared operational picture

The next step is a clearer handover between roles: field observations, planning assumptions and command decisions should stay attached to the same incident view.

Coordination that stays accountable

Future collaboration can add team notes, role-based reviews and a traceable decision history while keeping the incident commander in control.

The practical direction is collaboration, not automation for its own sake: richer field observations, clearer shift handovers and shared plan reviews can make an exercise more useful without making unsupported claims about real-world prediction.

THE INPUTS

Five scenarios, three ecological regions

Each scenario keeps its own ignition, weather and units. Switching simulations pauses the previous one without losing anything.

  • Replay

    Landiras I

    Gironde · July 2022

    Landes de Gascogne · maritime pine

  • Replay

    Saumos

    Gironde · July 2026

    Landes de Gascogne · maritime pine

  • Exercise

    Étoile massif

    Provence · mistral

    Provence limestone · garrigue

  • Replay

    Bug Fire

    California · August 2026

    Cismontane chaparral

  • Training

    Blank simulation

    Free starting point

    Set by the framing

A firefighter opening a control line with a drip torch across a field.
Control lines built during a run are cumulative and persistent in the engine.
Satellite view of smoke plumes drifting over a forested massif.

OPERATIONAL ACCESS

Open the console

Create an account, open the map, then let a WebMCP-capable agent work in the same tab. A new account opens the console on a five-step tour, which you can skip and replay at any time from the help menu.

  • Tactical map in 2D, 3D and globe
  • Adjustable weather, or a real hourly series
  • Three strategies compared by the engine
  • A log of the agent's calls