Built on Omnara

How it works

Every chase is driven live by its own AI agent running on Omnara. Here's what happens after you hit Start the chase.

Youstrategy, start, driver Driver agentone per chase, on Omnara Game serversimulates San Francisco Your screenplays the chase live plan drive call next intersection live
Youstrategy, start, driver
↓ plan
Driver agentone per chase, on Omnara
⇅ drive call · next intersection
Game serversimulates San Francisco
↓ live
Your screenplays the chase live
  1. 1

    You write the plan

    Up to 280 characters of strategy, a starting point anywhere from Lombard Street to Twin Peaks, and the AI model that drives. A moderation check screens your plan before it runs.

  2. 2

    An agent gets in the car

    Omnara starts a fresh agent for your chase from a driver profile: the model you picked plus one custom tool, drive. Your strategy is the agent's first message.

  3. 3

    Every intersection is a tool call

    Our game server simulates San Francisco's real streets and hills, with police hunting you. At each intersection it describes the scene: where the cops are, which exits you have, and what's closed. The agent calls drive with an exit, a speed and a line of dialogue, and Omnara holds that call until the server answers with the next intersection. The game waits up to 8 seconds for each decision. After that, the car keeps going straight.

  4. 4

    You watch it live

    The server streams every move to your browser, which plays the chase at real speed while the agent keeps driving ahead. Every line the driver says passes the same moderation check before it reaches your screen.

  5. 5

    Busted

    The chase ends when a patrol car catches you, or at 6:00 when the city locks down. Your time goes on the leaderboard. Run it back sends the same agent out again with the same plan, up to three runs in total, and it remembers how it got caught.

How the cops work

Build your own

The whole driver is one Omnara agent profile with a single custom tool. Our server answers each drive call through Omnara's API, so the model only ever sees the scene in words.

tools:
  drive:
    type: custom
    description: Make your decision for the coming intersection.
      Returns the next situation, or BUSTED when the chase is over.
    input_schema:
      type: object
      properties:
        exit:  { type: string, description: Exit letter from the list }
        speed: { type: string, enum: [floor_it, cruise, pull_over] }
        special: { type: string, enum: [none, jump] }
        say:   { type: string, description: One short in-character line }
      required: [exit, speed, say]