SiteKit

SiteKit is a proposed developer tool for websites that are useful, authorized, and not yet API-ready.

It records how a logged-in human uses a site, discovers the hidden HTTP or GraphQL interface behind the UI, and turns that knowledge into versioned local packages. The generated packages can run in data pipelines, cron jobs, and internal services. The same operations can also be exposed as MCP tools for agents.

SiteKit stops at the data interface layer. Sync jobs, analytics products, approval workflows, and domain-specific applications are built above SiteKit.

Discovery and execution are separate choices. A site can be discovered by a human recording or an autonomous browser agent. A final operation can execute through direct HTTP, Playwright browser automation, or a hybrid of both.

The LLM is in the authoring and repair loop. It is not in the production request path.

LLM provider credentials are needed only for discovery, review, and repair. Generated clients, MCP servers, auth handling, tests, and cron jobs run without model keys.

The Product Promise

sitekit onboard playmetrics \
  --url https://app.playmetrics.com \
  --goal "Expose roster, calendar, fees, fields, and safe game update operations."

SiteKit opens an isolated browser profile, records authorized API traffic while you click around, drafts an adapter, asks targeted review questions, and generates clients:

sitekit generate playmetrics --target typescript --target python
sitekit mcp serve playmetrics

What Gets Generated

adapters/playmetrics/
  adapter.yaml
  docs.md
  fixtures/redacted/
  schemas/
  tests/

packages/playmetrics-ts/
  src/client.ts
  src/raw.ts
  src/schemas.ts
  src/mcp.ts

packages/playmetrics-py/
  playmetrics/client.py
  playmetrics/raw.py
  playmetrics/models.py

Runtime Boundary

• Deterministic runtime: auth lifecycle, cookies, CSRF, pagination, retries, validation, rate limits, negative caching, and generated clients.
• LLM-assisted authoring: traffic interpretation, operation naming, schema descriptions, transform drafts, test suggestions, and adapter repair.
• Human review: write semantics, ambiguous business rules, safety gates, and known facts that cannot be inferred from traffic.

Discovery And Execution

Discovery mode:
  human       Record a logged-in walkthrough.
  autonomous Let an LLM explore within guarded browser limits.

Execution backend:
  http        Call discovered HTTP or GraphQL endpoints directly.
  browser     Drive the UI with Playwright.
  hybrid      Use the browser for setup, then call HTTP endpoints.

First Workflow