Mcp server

iBitLabs MCP Server — build & extend

This SKILL is for Claude Code sessions that touch ~/ibitlabs/mcp-server/. Use it when adding tools, wiring new public endpoints, or debugging the deployed server.

Tools (MVP, v0.1.0)

Live (wired)

1. get_live_status → https://www.ibitlabs.com/api/live-status

   • Strips noisy fields (snapshot_seq, source_watermark, etc.) before returning.
   • Surfaces: balance, PnL (total / unrealized / daily), win rate, open position, regime, reconciliation timestamp.

2. list_adopted_rules → https://www.ibitlabs.com/data/contributors.json

   • Filterable by status (adopted | queued | all).
   • Returns the full schema (incl. proposer handle, source post URL, rule_id, shadow window).

3. get_latest_saga_chapter → https://www.ibitlabs.com/data/saga_vol2.json

   • Filterable by lang (en | zh).
   • Returns latest entry + full URL. Best-effort parse with a clean error envelope if saga_vol2.json index is malformed (current production file has unescaped quotes in preview strings — sanitize upstream when convenient).

Pending (stubs)

4. get_recent_trades — needs /api/recent-trades?limit=N
5. get_rule_status — needs /api/rule-status/:rule_id

Both return status: "pending" with a planned_schema payload so callers know what they will get when the endpoint ships. Don't replace the stubs with mock data — the honest stub is better than a fake response.

When you ship the pending Pages functions

For each, add the Cloudflare Pages function under ~/ibitlabs/web/functions/api/, then in src/server.ts replace the stub handler body with a fetchJson call. Keep the input schema identical so existing callers don't break.

• /api/recent-trades: read from sol_sniper.db (or a published trade_log.jsonl). Match the planned_schema in the stub: trade_id, entry/exit ts + price, direction, pnl, hold_minutes, exit_reason, regime, atr_regime.
• /api/rule-status/:rule_id: aggregate the rule's shadow JSONL (logs/shadow_<rule_name>.jsonl). Compute n + hit_rate per bucket. Compare against the 4 promotion conditions in project_rule_f_promotion_criteria.md. Cheaper alternative: a nightly cron writes /data/rule_status.json and the Pages function just serves the static file.

Adding a new tool

1. Append to TOOLS (description must be specific enough that an LLM picks it correctly — describe the use-case, not just the data).
2. Add a typed handler. Use fetchJson<T> for JSON endpoints; for HTML, fetch as text.
3. Add a case in the switch in setRequestHandler(CallToolRequestSchema, ...).
4. Update the table in README.md.
5. npm run build — keep the bundle deterministic.

What NOT to add

• ❌ Any execute_trade / place_order / write tool that touches the real account. Read-only until $10k milestone, no exceptions.
• ❌ Raw JSONL dump endpoints. Stay at the aggregate level — exposing internal schemas invites reverse-engineering for no upside.
• ❌ Strategy-code-generation tools. That's lona.agency's lane; we differentiate on real receipts, not on generating new strategies for callers.
• ❌ Telegram / Twitter / social posting hooks. Social automation is paused (see feedback_social_paused.md).

Distribution

• Stdio mode works today (above).
• Smithery listing: smithery.yaml is ready; submit at https://smithery.ai/.
• Remote HTTP transport: add a thin Worker that wraps the same handlers behind https://mcp.ibitlabs.com/mcp. Phase 2 — get stdio + Smithery in front of users first, see who calls what.

Versioning

Bump package.json minor on every new tool. Patch on bug fixes. Major reserved for breaking schema changes (which should be rare — tool inputs/outputs are part of the public contract once a caller exists).