Outpost
A daily browser roguelike where four AI advisors run a fragile colony, argue about what to do next, hold grudges, and sometimes try to kill you.
You play the leader. Each day you get one run. It lasts 8–12 turns, ends when the colony holds out, collapses, ousts you or kills you, and produces a post-mortem timeline you can share.
Status (October 2026): playable at outpost-ai-roguelike.vercel.app with Google sign-in, one run a day. That's the turn-based version this README describes. It works, but my first full run in production showed it isn't much of a game yet: my choices felt disconnected from the numbers, and reading four advisors argue got old fast. So I'm rebuilding it as a short real-time holdout on a new deterministic engine. The current version stays up until the new one is playable end to end. See Where it's going.
Why this exists
I wanted to find out whether a small model voicing four advisors, given personalities and grievances and a shared world to mutate through tool calls, could produce stories that feel like a person wrote them. Not "the AI wrote a story." Stories where you can point at turn six and say that's the moment Marisol stopped trusting me, and that's why she sided with Kel later.
Most LLM-driven games either let the model freely narrate (fun, but the world doesn't actually change) or pin everything down so tightly the model is just a flavor layer. Outpost picks a middle path: a deterministic engine owns state, the agents propose changes only through validated tool calls, and a single batched debate call decides the turn. The debate is one forced structured tool call that returns every advisor's reaction plus three options, so four advisors cost one LLM call, not four.
Some of it works. Some of it I would do differently next time. It's all in docs/PLAN.md's "Errors & Learnings" log.
What's in the box
- Four advisors with three personality axes (Order/Liberty, Collective/Individual, Cautious/Bold) and persistent memory across turns. They form alliances. They die. Dead means dead.
- Three interlocking systems: a Grievance Stack (advisors track every choice you make against them), a Delayed Consequence log (decisions seed events 1–3 turns later), and a Legitimacy meter (drops too far and the colony stops obeying).
- Three launch settings: arctic research station, orbital colony, desert outpost. Each is a YAML schema, so adding a fourth setting is a config change, not code.
- Permadeath end-states: the colony endures, collapses, ousts you, or kills you. The post-mortem timeline shows the chain of decisions that got you there, plus the consequences you never set off and the coups that nearly happened.
- One run per account per day, server-authoritative. Resets at UTC midnight.
- Shareable post-mortem with an OG image card sized for Twitter, Bluesky, and Discord.
Inspirations: Papers Please, Don't Starve, Detroit: Become Human, Rimworld, Frostpunk.
Stack
- Backend: FastAPI (Python 3.12), LangGraph for the turn graph, and the raw Anthropic SDK pointed at any Anthropic-compatible endpoint. Production runs DeepSeek V4.1 Flash through OpenRouter, prompt-cached, with the debate as one forced tool call. Deployed on Railway.
- Frontend: Next.js 16 App Router, TypeScript strict, Tailwind v4, shadcn/ui re-themed for pixel-art, DiceBear for portraits. Deployed on Vercel.
- Data: Supabase Postgres via SQLAlchemy async + psycopg3. Supabase Auth (Google OAuth) with JWTs validated locally in FastAPI.
- Observability: every LLM call's model, tokens and billed cost are stored with its turn. Langfuse traces the LangGraph steps. LangSmith
agentevals runs the merge-gate evals (personality consistency, tool-call validity, narrative repetition, decision coherence, game balance).
Why these choices? docs/research/ has the long version. The short one: the raw SDK keeps cache and tool-call control honest, LangGraph gives durable state across the SSE pause/resume, and Supabase is the cheapest way to get OAuth and Postgres in one place.
The model was picked blind. Twenty models played 12 real turns, the game's own validator threw out broken output, and four Opus judges scored the rest without knowing which model wrote what. The finalists then played full games. DeepSeek V4.1 Flash matched the previous model's quality at 1.41 cents a game in real spend, down from about 6.
Repo layout
Outpost/
├── docs/ SPEC, ARCHITECTURE, PLAN, research/
├── backend/ FastAPI app
│ ├── app/ game engine, agent graph, auth, eval
│ ├── tests/ unit + integration + eval (markered)
│ ├── alembic/ ORM migrations (public schema)
│ └── supabase/migrations/ CLI migrations (auth schema, RLS)
├── frontend/ Next.js app
│ ├── app/ (public) + (authed) route groups
│ ├── components/ dashboard panels, post-mortem timeline
│ └── hooks/ SSE consumption
├── .github/workflows/ test, eval, deploy
├── CLAUDE.md conventions for AI-assisted development
└── README.md you are here
Local development
Prerequisites: Docker, Node 22+, Python 3.12+, uv, Supabase CLI, Railway CLI (for deploy), gh.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
