Shipwright

The AI SaaS starter kit where the AI part is actually finished.

Most boilerplates bolt a chat call onto an auth template. Shipwright ships the hard parts of an AI product — streaming, tool use, structured outputs, prompt versioning with evals, and per-user token metering wired for billing.

Try the demo chat Read the code

Streaming done right

SSE chat endpoint with token-by-token streaming, client disconnect handling, and a final usage event — not a toy fetch wrapper.

src/lib/ai/stream.ts

Token metering & cost attribution

Every request is recorded per user with exact cost in micro-USD, including prompt-cache reads and writes. Quota checks and usage-based billing build on one table.

src/lib/metering/meter.ts

Typed tool use

Declare tool inputs once with Zod: JSON Schema for the API, runtime validation with model-readable errors, and typed execute() — plus a production agent loop with iteration caps.

src/lib/ai/tools.ts

Structured outputs

Schema-guaranteed JSON via the API's structured outputs — no "respond only with JSON" prompting, no parse-retry loops.

src/lib/ai/generate.ts

Versioned prompts + evals

Prompts live in a versioned registry, and an eval harness with deterministic graders and an LLM judge gates prompt changes in CI.

src/lib/ai/evals/runner.ts

Rate limiting

Per-caller limits on the AI endpoints out of the box, with a seam to swap in Redis for multi-instance deploys.

src/lib/ratelimit.ts