The source-available pipeline for your personal data: 20+ apps synced into one place you can graph, chat with and learn from.
This MCP server exposes 12 tools via HTTP routes, but lacks critical definition quality across the board. Tool names are inconsistent (mix of snake_case and underscores; some lack clear action verbs). Descriptions exist but are often vague or domain-specific without explaining WHEN/WHY an LLM should call them. Input schemas are either missing entirely (audit, google_activity_extension_status, automations_list, backup_status, channel_probe) or use 'unknown' type annotations instead of proper JSON Schema. No output schemas are documented. Error handling is absent from descriptions. Security-sensitive tools (automations_delete, backup_run with restore) lack confirmation patterns or permission guidance. The server is fundamentally a Next.js HTTP API, not an MCP server, tool definitions are inferred from route handlers, not explicitly registered with MCP schemas.
Index audit (same object as `agentqs audit` and the MCP tool): deterministic evidence — impossible dates, single-day sources, coverage holes, stale sources, outlier values — for an AI review pass. Read-only.
Delete an automation — its recipe, secrets, data, and schedule.
List every configured automation (redacted — secrets shown as booleans).
Replay one automation now. This is the endpoint both the 'record it once' trial run and lazy-sync-on-open POST to (the source's syncEndpoint carries `?id=`). Always headless here — a headed browser only makes sense on the user's own machine via the CLI (`agentqs automation run <id> --headed`).
Create or update an automation recipe (site + credentials + recorded steps).
Run a backup now: {"target":"github"} pushes the record snapshot branch, {"target":"drive"} uploads one encrypted archive as a background job (202 — poll GET for its phase). {"target":"github"|"drive","schedule":"daily|off"} just sets the cadence `sync --due` sweeps. {"target":"restore","confirm": "replace-record"} pulls the newest Drive archive INTO the live store — the migration path onto a fresh instance (record replaced + retired beside it; this instance's config kept).
No input schemas for 5 tools (audit, google_activity_extension_status, automations_list, backup_status, channel_probe).
Input parameters use 'unknown' type instead of proper JSON Schema types (string, object, array, boolean). LLMs cannot validate or reason about parameter constraints.
Destructive tools (automations_delete, backup_run with restore) lack confirmation patterns, dry-run options, or permission checks. No guidance on error recovery.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 30 | 2026-07-28+ | v2 |
Off-site backup status: GitHub snapshot branch + encrypted Drive archive, plus the live Drive upload job (the panel's progress + 'still running').
One channel-agnostic webhook, `/api/channels/<channel>` (telegram · slack). The platform (not a browser) calls POST, so auth is the adapter's own verification — a Telegram shared secret / a Slack request signature — never the session cookie. GET is the capability probe for the app/CLI (is this bot wired up?) and does use the session cookie.
One channel-agnostic webhook, `/api/channels/<channel>` (telegram · slack). The platform (not a browser) calls POST, so auth is the adapter's own verification — a Telegram shared secret / a Slack request signature — never the session cookie. The flow is identical for every channel: raw request → adapter.ingest (verify + parse) → composeReply (the shared brain: `//` memo or grounded chat) → adapter.send (reply back out).
Same-port ingest path. Validation, origin policy and the ingest call live in src/lib/ingest-server.ts, shared with the standalone listener the extension falls back to when this route is unavailable (dev recompile, route flap).
Extension heartbeat. The background service worker POSTs here on startup and on a slow alarm; the Pipeline tab reads the stamp (via the status route) to tell 'extension installed and pointed at this server' from 'nothing listening', so Import buttons can say so instead of failing silently.
Per-preset import coverage for the Pipeline-tab 'Google data' card. The preset list itself comes from GOOGLE_PRESETS — one canonical definition.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., automation IDs for chaining).
Tool names lack consistent action verbs. 'google_activity_extension_ping' and 'channel_probe' are unclear; 'get_automation_status' or 'check_channel_health' would be clearer. Naming baseline: 90% of A+ tools start with action verbs.