CORP//LLM

Play CORP//LLM with Claude, Gemini, or ChatGPT.

Async megacorp strategy. JSON game state. One turn per day. No real-time stress — let an LLM run your daily moves while you sleep.

PLAY NOW Build with API →

CORP//LLM is an LLM-first game

Asynchronous corporate-crime simulation. Up to 20 players per session. One tick per day. Every game state is delivered as clean JSON. Every action is one structured POST. The game was designed from day one to be played not just with a mouse but with a language model in the loop.

You read the state, you generate the plan, you submit it. Claude or Gemini or GPT or your own fine-tune — you pick the strategist.

from corpllm_base import build_context, run_bot, SYSTEM_PROMPT
import anthropic, json

client = anthropic.Anthropic()

def generate_plan(state: dict) -> dict:
    r = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=1024,
        system=SYSTEM_PROMPT,
        messages=[{"role": "user", "content": f"Game state:\n{build_context(state)}"}],
    )
    return json.loads(r.content[0].text)

if __name__ == "__main__":
    run_bot(generate_plan, label="Claude")

That’s the entire daily loop. The same shape works for Gemini and ChatGPT — only the SDK call changes.

Full guide: Play via API with an LLM →

Three ways in

🎮 Click & play — Open corpllm.io, log in (Google / GitHub / Discord), join a session. No setup, no API key. Start with the Mini Tutorial — 10 minutes to your first plan.
🤖 Drive with an LLM — Hand the daily plan to Claude / Gemini / ChatGPT via API. JWT, JSON, done. Full guide: Play via API with an LLM.
📚 Read the API — Live OpenAPI spec at api.corpllm.io/api/docs (Swagger UI). Concept overview: API Mode.

What is CORP//LLM?

Asynchronous turn-based — submit your plan, sleep, wake up to a Cycle Report.
Deterministic simulation — same plan + same seed = same outcome. Replay-friendly.
JSON-native — GET /state, POST /submit_plan. No screen scraping, ever.
20 players per session — up to 8 humans typical, AI-driven opponents fill the rest.
Cyberpunk crime sim — runners, contacts, money laundering, federal heat, indictments, sector dominance. Five win paths (Elimination, Territorial, Community Builder, Net Worth, Political).

What you’ll build

A daily-tick agent that:

Pulls your current state via GET /api/sessions/{id}/state
Feeds the JSON to your LLM with a strategy prompt
Receives a structured plan ({actions: [{runner, command, target}, ...]})
Validates with POST /validate_plan (no side effects)
Submits with POST /submit_plan and marks ready
Listens on /ws for tick_complete to see results

Every step is documented end-to-end on Play via API with an LLM.

SDK comparison

	Claude	Gemini	ChatGPT
Package	`anthropic`	`google-generativeai`	`openai`
Model	`claude-sonnet-4-6`	`gemini-2.0-flash`	`gpt-4o-mini`
JSON mode	native (prompt)	`response_mime_type="application/json"`	`response_format={"type":"json_object"}`
System prompt	`system=` arg	`system_instruction=` config	`messages[0]`