Waggle

An MCP server for planning and overseeing the tecture-graph project from the outside. It gives coding agents a persistent, queryable place to:

• Log tecture-graph researches — what is being investigated, why, and what came of it
• Log development activities — each step performed, linked to a research or directly on the code
• Publish test results — one record per test run, optionally linked to a research
• Write and read overall project progress — an append-only journal of where the project stands

Waggle owns its own SQLite database and knows nothing about the project it tracks; the tracked project registers Waggle as just another MCP server. (Named after the waggle dance — how bees report their findings back to the hive. Sibling of mcp-bumble, whose architecture — including the OAuth 2.1 HTTP transport — it follows.)

Tools

The research tools carry the tecture_ prefix deliberately, so they don't collide with other research-log connectors in the same Claude workspace.

Publishing full Vitest reports (with per-test console logs)

Vitest's built-in JSON reporter (--reporter=json) includes per-test names, statuses, durations and failure messages — but no console output. To capture logs per test, use the bundled custom reporter, which listens to onUserConsoleLog (its taskId matches TestCase.id) and writes a payload that can be passed to publish_test_results verbatim:

npm run test:report   # = vitest run --reporter=default --reporter=./scripts/vitest-waggle-reporter.mjs
# → waggle-report.json: { suite, durationMs, tests: [{ name, file, status, durationMs, error?, logs? }] }

WAGGLE_SUITE and WAGGLE_REPORT_FILE override the suite name and output path. The reporter is self-contained — copy scripts/vitest-waggle-reporter.mjs into any Vitest project that publishes to Waggle.

Dashboard

A read-only web interface over the same database — browse researches, activities, test runs and the progress journal without going through MCP:

OAUTH_ADMIN_PASSWORD=... npm run ui          # http://127.0.0.1:3204

It is protected by the same admin password as the MCP server (OAUTH_ADMIN_PASSWORD, the one the OAuth consent page uses) and refuses to start without it. Logging in sets an HttpOnly session cookie (7 days, kept in memory — a restart signs everyone out). The dashboard never mutates data.

In HTTP mode (WAGGLE_TRANSPORT=http) the dashboard is also served from the MCP server's own origin — every path the MCP/OAuth routes don't claim — so the hosted instance's dashboard lives at https://waggle.heycasper.uk/ behind the same login. No separate process or port needed.

Transports

STDIO (default) — for local Claude Code / Claude Desktop:

{
  "mcpServers": {
    "waggle": { "command": "node", "args": ["/path/to/mcp-waggle/dist/index.js"] }
  }
}

HTTP + OAuth 2.1 (WAGGLE_TRANSPORT=http) — for use as a claude.ai custom connector. Streamable HTTP at /mcp, OAuth 2.1 with dynamic client registration, PKCE, a password-gated consent page, opaque 15-minute access tokens and rotated 90-day refresh tokens (stored hashed in a JSON file), and RFC 8707 resource binding. Same design as mcp-bumble v1.1.

Setup

npm ci
npm run build       # bundles to dist/ and copies migrations
npm run smoke       # end-to-end STDIO smoke test of the built server

node dist/index.js migrate applies migrations and exits (they also run automatically on startup).

Deploy

The hosted instance runs on the Raspberry Pi as mcp-waggle.service (checkout at /home/pi/mcp-waggle, config in /etc/mcp-waggle.env, deployed from origin/main):

npm run deploy      # = ./scripts/deploy-pi.sh

The script verifies the local tree is clean and pushed, runs the test suite (skip with SKIP_TESTS=1), then over SSH: syncs the Pi checkout to origin/main, npm ci, builds, restarts the service, and health-checks it locally and via https://waggle.heycasper.uk. Override PI_HOST, PI_DIR, SERVICE, BRANCH, or PUBLIC_URL via env if the setup moves. npm run db:seed (= node dist/index.js seed) inserts a snapshot of real tecture-graph tracking data (captured 2026-07-09 from the live Waggle instance: the CodeGraph DB-survey research with its linked activities and test run, plus the progress journal) into DB_PATH; it is idempotent, so re-running never duplicates rows.

Development

• Stack: TypeScript, Node ≥ 20, @modelcontextprotocol/sdk, Express 5 (HTTP transport), SQLite via better-sqlite3, Drizzle ORM with committed migrations, tsup, Vitest, ESLint.
• Layout: src/index.ts (entry/dispatch) → src/transport/{stdio,http}.ts → src/server.ts (tool registration) → src/tools/* (pure functions + thin MCP wrappers) → src/oauth/* (provider, JSON-file token store, consent page) → src/ui/* (read-only local dashboard: Express routes + server-rendered HTML) → src/db/* (schema, connection + migration runner, seed fixtures) → src/lib/ids.ts (nanoid prefixes: res_, act_, run_, prog_).

npm test               # all Vitest suites (fresh in-memory SQLite per test)
npm run test:coverage  # v8 coverage, 80% thresholds enforced
npm run lint
npm run typecheck
npm run db:generate    # regenerate migrations after editing src/db/schema.ts