brink

A single-player, cinematic geopolitical roleplay: you are the supreme executive of a great power, handling only the geopolitical brief — taking calls from world leaders, hearing your advisors, making decisions — in a persistent world grounded in current real-world geopolitics.

Built as a fork of DMCP (LLM-as-DM, MCP server as persistent state + tools), with a geopolitical rules core ported from the BalanceOfTerror engine.

To play it, read docs/MANUAL.md.

The design brief is REBOOT-DESIGN-AND-REQUIREMENTS.md and it is the authority on everything below. Read it first if you're working on the code.

The shape of it

• The LLM is the Game-Master: it classifies a fuzzy action into which mechanic fires, narrates, and calls tools. It never adjudicates.

• The MCP server is referee + ledger: authoritative resolution, persistence, and the handful of rules the GM cannot talk its way out of.

Three hard rules that constrain every change here

1. Show, don't tell. If the player-facing surface ever states a disposition or a raw value — DOCTRINE: HAWK, cohesion 80.0, (DOVE advisor) — that is a bug, not a style preference. The player deduces state from behavior, dialogue, and consequence.

2. Resolve, then narrate. The GM never narrates an outcome before the server returns it. It calls the resolve tool, receives an authoritative outcome payload carrying an explicit narration constraint, and narrates from that payload. The fiction can be florid about the how; it can never lie about the what.

3. No credential in the working tree. The GM runs keyless, on a Claude subscription (OAuth). There is no standing pay-per-token API key anywhere in this repo, and development works against a mock GM backend. See §2.4 of the brief — this is a requirement, not guidance.

Related MCP server: DMCP

Fork discipline (§10 of the brief)

Changes sort into three buckets, and generic mechanism must never share a commit with geopolitical content. That separation is what keeps the offer-back-upstream workflow mechanical instead of archaeological.

The line: "a resource can be declared conserved" is a DMCP feature; "prestige is conserved across three great powers" is our game.

docs/IMPROVEMENTS-TO-DMCP.md is the full account of what changed in DMCP and which parts are generic enough to lift — written for DMCP users deciding whether any of it is worth taking, as much as for upstream.

Branch convention

• upstream remote tracks DMCP. git fetch upstream && git merge upstream/main pulls their changes.

• Work intended for upstream branches from upstream/main as offer/<topic>, touches only DMCP-generic files, and carries no geopolitical content — so it cherry-picks cleanly. (offer/, not upstream/: a local branch named upstream/foo collides with the upstream remote's own ref namespace, and git will happily let you create the ambiguity.)

• Everything else branches from main.

• Upstream's own README is preserved in history: git show upstream/main:README.md (setup instructions and the full ~170-tool inventory).

Development

npm ci               # install
npm run build        # tsc + client build
npm run dev          # run the server
npm run test:run     # vitest
npm run lint         # eslint      (CI gate)
npm run typecheck    # tsc --noEmit (CI gate)

Local MCP testing: npx @modelcontextprotocol/inspector node dist/index.js

Write lots of tests. DMCP arrived with essentially no coverage of business logic, so there is no upstream safety net under the invariants this game depends on. TDD is mandatory — failing test first. See CLAUDE.md for the test harness and the codebase gotchas worth knowing before you start.

Status

Tranche 1 is complete. The text-first vertical slice runs:

npm run build:server        # once
npm run play -- --real      # Claude as the game-master
npm run play                # placeholder game-master, spends nothing

Graphics amplify a loop that already works; they cannot rescue one that doesn't — so this had to exist before anything cinematic.

What's in place:

• The referee. All three enforced rules from §4.2 have real mechanisms: prestige is a conserved constraint set moved only through an atomic transfer; DEFCON is bounded and resolve_only, so it has exactly one writer; declared on-expiry consequences are applied by the server itself, so a countdown lands regardless of the fiction.

• The rules core, ported from the BalanceOfTerror Python engine as pure functions with injected RNG, verified against that engine's own test suite.

• The resolve path, which makes resolve-then-narrate enforceable: the payload carries an explicit narration constraint, and tests re-read the database after every call to prove prose and ledger cannot disagree.

• The GM layer: system prompt, persona seeds whose dispositions exist only as behavior, a GMBackend seam with a deterministic offline mock, and the real game-master — Claude Code driven headlessly under a subscription, with no credential anywhere in the tree.

Verified live: the GM streams an outcome-neutral reaction, calls the referee itself, and narrates only after the payload lands. DEFCON moved 5→4 with a matching audit row — and since DEFCON carries a resolve_only constraint, that write had no other route available to it.

Tracked in the issues.

Contributing, and security

• CONTRIBUTING.md — how to get set up, and the two rules that catch people out.

• SECURITY.md — how to report something, what leaves your machine, and why there is no API key in this repository.

• docs/PUBLISHING.md — the repository topology, and the two routes work takes out of here: a tree copy to the public repo, and offer/* branches to DMCP.

License

MIT. Copyright is shared: DMCP's original notice is retained alongside brink's own — see LICENSE.