MCP Registry
⚠️ DRIFT ALERT: CLAUDE.md says 8 MCP servers — live probe shows 22 entries, MCP-INVENTORY.md documents 21
This is a known, documented drift. The discrepancy has three layers:
-
CLAUDE.md §Memory & Tools states: “MCP servers active (8): openclaw-tools ✓, aurora-orchestrator ✓ (20 tools), context7 ✓, playwright ✓, shortio ✓, 21st-magic ✓, dune ✓ (HTTP), figma ✗ (broken)” — this was accurate when written but has become stale as 13 additional MCPs were added to
~/.mcp.json. -
MCP-INVENTORY.md (
[[workspace/MCP-INVENTORY]]) documents 21 MCPs (added in vendor audit 2026-05-03), organized as: 8 original + 13 additional. This is the current canonical inventory. -
Live probe (
claude mcp list 2>/dev/null | wc -l) returned 22 lines (includes 1 header/separator line + 21 servers + blank = ~22 raw). Active connections: 14 connected, 6 failed, 1 HTTP (dune) = 22 entries total.
Action required: Update CLAUDE.md §Memory & Tools to say “MCP servers (21): see MCP-INVENTORY for full list.” This is a G-KB-SYNC-WITH-CLAUDEMD adjacent finding (not a KB dir, but same principle: CLAUDE.md must match live state).
MOC for all Model Context Protocol servers registered in ~/.mcp.json. Covers 21 servers across 8 internal/critical + 13 extended tiers. Read this hub when selecting the right MCP for a task, debugging a failed connection, or planning new MCP additions. The canonical config is ~/.mcp.json; the detailed inventory is [[workspace/MCP-INVENTORY]]. Owner: Henry Hill / Claude Opus 4.7.
Quick reference
| Field | Value |
|---|---|
| Canonical config | ~/.mcp.json |
| Canonical inventory | [[workspace/MCP-INVENTORY]] |
| Live state probe | claude mcp list 2>/dev/null | wc -l |
| Documented count | 21 (MCP-INVENTORY.md source) |
| Live count | 22 lines returned (21 servers + formatting) |
| CLAUDE.md count | 8 (stale — needs update) |
| Drift | CLAUDE.md 8 vs live 21 — update CLAUDE.md |
| Connected (working) | 14 of 21 (probed: 2026-05-03) |
| Failed to connect | 6 of 21 (see Tier B below) |
| Last audit | 2026-05-03 |
| Drift alert | manual (no automated MCP drift check yet) |
| Governance gate | G-NO-PLAINTEXT-CREDS (all creds via env vars in ~/.mcp.json) |
| Owner agent | system / claude-code |
Tier A — Original 8 (CLAUDE.md-documented, active session)
These are the 8 MCPs originally in CLAUDE.md. They are the primary tools for Claude Code sessions.
| # | MCP | Type | Live Status | Tools/Purpose |
|---|---|---|---|---|
| 1 | openclaw-tools | stdio Python | ✅ Connected | 16 tools: memory_search, memory_save, workspace_query, run_script, portkey_status, agent_status, gateway_logs, supabase_query, hubspot_lookup, memory_health, stack_query + semantic_search, web_research, library_docs, library_search, send_sms_compliantly |
| 2 | aurora-orchestrator | stdio Node | ✅ Connected | 20 Aurora tools: reply_to_imessage, reply_to_thread, react_to_imessage, hubspot_deal_create, hubspot_deal_update, hubspot_lookup, omni_contact_timeline, openphone_thread_history, semantic_search, memory_search, schedule_followup, escalate_slack, flag_for_review, post_final_summary, send_openphone_sms, supabase_query, docusign_contract_draft, dispatch_dispo_email, consult_acquisitions, consult_betterfiles, consult_dispositions, ask_clarifying |
| 3 | context7 | stdio npx | ✅ Connected | Live SDK + library documentation lookup |
| 4 | playwright | stdio npx | ✅ Connected | Headless browser automation |
| 5 | shortio | stdio npx | ✅ Connected | URL shortener ops via Short.io API |
| 6 | 21st-magic | stdio npx | ✅ Connected | React component browsing + import |
| 7 | dune | HTTP | ✅ Connected | Blockchain analytics via Dune API (HTTP MCP, not stdio) |
| 8 | figma | stdio npx | ✗ Failed | Design file inspection — known broken per CLAUDE.md (Phase 11.3a fix pending) |
Tier B — 13 Extended MCPs (added in vendor audit 2026-05-03)
These were added to ~/.mcp.json as part of the vendor deep-audit. Some need credential setup.
| # | MCP | Live Status | Creds needed | Purpose |
|---|---|---|---|---|
| 9 | postgres-diag | ✗ Failed to connect | POSTGRES_MCP_DIAG_URL (check env) | Query plan analysis, slow-query ID, index recommendations |
| 10 | firecrawl | ✅ Connected | FIRECRAWL_API_KEY | Web scraping/crawling/mapping → markdown |
| 11 | sequential-thinking | ✅ Connected | none | Multi-step structured reasoning trace |
| 12 | memory | ✅ Connected | none (file-backed) | Cross-session knowledge graph (different from per-agent SQLite) |
| 13 | github | ✅ Connected | GITHUB_TOKEN | Repos / PRs / issues / Actions |
| 14 | supabase | ✅ Connected | SUPABASE_ACCESS_TOKEN | Supabase admin (read-only) |
| 15 | hubspot | ✗ Failed to connect | HUBSPOT_PAT (check env) | CRM deal/contact/company CRUD |
| 16 | slack | ✗ Failed to connect | SLACK_TEAM_ID missing | Channel/message ops — ⚠️ DEPRECATED npm, SLACK_TEAM_ID not set |
| 17 | stripe | ✗ Failed to connect | STRIPE_SECRET_KEY (check env) | Payments / subscriptions |
| 18 | notion | ✅ Connected | NOTION_TOKEN | Pages/databases |
| 19 | google-workspace | ✗ Failed to connect | OAuth refresh token (henry@reri.co was wired 2026-05-03 — re-verify) | Gmail/Drive/Calendar/Docs/Sheets |
| 20 | discord | ✅ Connected | DISCORD_BOT_TOKEN | Channel posts, message ops |
| 21 | airtable | ✅ Connected | AIRTABLE_API_KEY | Database ops |
| 22 | perplexity | ✅ Connected | PERPLEXITY_API_KEY | Sonar API research (4 modes: fact/research/deep/reason) |
Note:
perplexityappears in live probe but was listed as #5 in CLAUDE.md §Memory & Tools sometimes — it is MCP #22 in this count (beyond MCP-INVENTORY’s 21). Live probe confirmed 22 lines;perplexityis the 22nd server.
Live state snapshot (2026-05-03)
| Metric | Documented | Live (probed) | Drift | Status |
|---|---|---|---|---|
| CLAUDE.md count | 8 | 22 | +14 | ⚠️ CLAUDE.md stale |
| MCP-INVENTORY.md count | 21 | 22 | +1 | perplexity not counted in v1 inventory |
| Connected (working) | 21 | 14 | −7 | 6 failed + figma known broken |
| Failed — known reason | 1 (figma) | 6 | — | figma + slack(TEAM_ID) + postgres-diag + hubspot + stripe + google-workspace |
| Cred-proxy eligible | pending | 0 deployed | — | awaiting B1-B3 ratification nemoclaw-audit-2026-05-03 |
Failed MCP diagnosis (2026-05-03)
| MCP | Failure reason | Fix |
|---|---|---|
figma | Known broken (Phase 11.3a) | Fix scheduled per openclaw-fragmentation-fix-2026-05-01 |
postgres-diag | POSTGRES_MCP_DIAG_URL env may not be set in session | Verify in ~/.openclaw/master.env |
hubspot | HUBSPOT_PAT env may not be exported to MCP context | Verify export in ~/.mcp.json env block |
slack | SLACK_TEAM_ID missing + deprecated npm package | Find replacement package; add SLACK_TEAM_ID |
stripe | STRIPE_SECRET_KEY env may not be exported | Verify in ~/.mcp.json env block |
google-workspace | OAuth refresh token configured 2026-05-03 — may need session reload | Cmd/Ctrl+Shift+P → Developer: Reload Window |
When to use which MCP
For substrate / data queries
openclaw-toolssupabase_query— Supabase CCP production queries (73 tables)supabase— Supabase admin ops (auth, storage, realtime) via supabasepostgres-diag— slow query / index advice from pg_stat_statements (when working)openclaw-toolssemantic_search— cross-source RAG retrieval (messages + omni_events + KB + agent_memories)
For code / SDK / docs
context7— live library docs (preferred over training data for any SDK/framework)github— PR/issue/code search across repos includingtraewayrer/openclaw-vaultopenclaw-toolslibrary_docs— same Context7 via openclaw-tools wrapper (cost-tracked)
For web / external research
perplexity— synthesized research with citations (4 modes: sonar, sonar-pro, deep-research, sonar-reasoning)firecrawl— scrape/crawl/map external pages → markdown
For CRM + communications
hubspot— deal/contact/company CRUD via PAT (when working)aurora-orchestrator— 20 Aurora tools for deal dispatch, CRM ops, SMS, iMessagediscord— Discord channel posts via bot tokenairtable— database ops (now working per live probe)
For browser + design
playwright— headless browser (QA, form tests, screenshots)21st-magic— React component browsing and importfigma— design file inspection (when fixed)
For reasoning + memory
sequential-thinking— multi-step reasoning with structured tracememory— cross-session knowledge graph (complements per-agent SQLite DBs)openclaw-toolsmemory_search/memory_save— per-agent SQLite hybrid search (44 DBs)
For payments + business
stripe— payments/subscriptions/customers (when working)notion— pages/databases (now working per live probe)google-workspace— Gmail/Drive/Calendar/Docs
MCP config location
~/.mcp.json — master MCP configuration (21 servers)
All credentials referenced via environment variable names — no plaintext keys in ~/.mcp.json per G-NO-PLAINTEXT-CREDS. Actual values loaded from ~/.openclaw/master.env (chmod 600, 19 sections, 111+ vars).
After installing a new MCP or changing creds: Cmd/Ctrl+Shift+P → Developer: Reload Window to register in active session.
Internal MCPs to build (Tier D backlog from MCP-INVENTORY)
These high-ROI scripts would benefit from MCP wrappers (not yet built):
| Priority | MCP name | Wraps | LOC |
|---|---|---|---|
| HIGH | dispoBroadcast | dispo-blast-engine.js | 1209 |
| HIGH | createDealFromParsed | hubspot-deal-creator.js | 1006 |
| HIGH | checkCompliance | compliance-gate.js | 244 |
| HIGH | generateResponse | response-generator.js | 1737 |
| HIGH | sendSmsViaLine | quo-handler-enhanced.js | 2062 |
| HIGH | salesmessage-tools | salesmessage-handler-v4-complete.js | 4884 |
| MEDIUM | sendSmsCompliantly | dispo-send.js | 1008 |
OSIL / NemoClaw forward references
- NemoClaw cred-proxy pattern (B1-B3 pending ratification): if ratified, MCPs that require API keys would route via L7 credential injection proxy at
:18901rather than env-loading keys directly. This would make ALL 21 MCPscred-proxy-eligible: yes. See nemoclaw-audit-2026-05-03. - OSIL B10 (LiveKit): LiveKit creds added to 1Password Aurora vault. LiveKit agent scripts validated. No MCP yet (deferred to OSIL Phase 6). Forward-ref: openclaw-self-improvement-layer-2026-05-03.
Cross-links
Agents that touch this
- _summary — aurora-orchestrator MCP (20 tools) is Aurora’s primary interface
- _summary — uses openclaw-tools supabase_query + semantic_search
- _summary — uses context7 + github for research
- _summary — anthropic tools routed through portkey proxy
Skills that invoke this
- acquisitions-outreach — openclaw-tools: send_sms_compliantly, supabase_query
- dispo-blast — openclaw-tools: dispo_broadcast, supabase_query
- il-marketplace-pull — openclaw-tools: supabase_query + firecrawl (scraping)
- hubspot-deal-ingest — openclaw-tools: hubspot_lookup + supabase_query
- mcp-builder — for creating new MCPs
Plans that govern this
- vendor-deep-audit-comprehensive-2026-05-02 — surfaced the 13 additional MCPs
- openclaw-fragmentation-fix-2026-05-01 — figma fix (Phase 11.3a)
- nemoclaw-audit-2026-05-03 — cred-proxy pattern (B1-B3)
Feedback rules
- feedback_no_plaintext_creds — G-NO-PLAINTEXT-CREDS: env vars only in ~/.mcp.json
- feedback_kb_sync_with_claude_md — G-KB-SYNC-WITH-CLAUDEMD analog: CLAUDE.md must reflect live MCP count
- feedback_check_resources_before_asking — check MCP-INVENTORY before adding new MCPs
KB / source docs
- README — supabase MCP + supabase_query tool
- README — github MCP (repos, PRs, issues)
- README — openclaw-tools + aurora-orchestrator brain routing
System maps
- mcp-server-catalog — visual MCP catalog map
- request-lifecycle — how MCP calls flow through Portkey proxy
Related cluster — System catalog
- service-registry — systemd units that back some MCP servers
- port-registry — no direct port binding (stdio MCPs)
- portkey — LLM calls from agents flow through Portkey; MCPs are tools not routing
- supabase — supabase MCP targets CCP project (svueekfvfrvhylxygktb)
- github — github MCP targets traewayrer/openclaw-vault + other repos
- discord — discord MCP + discord skill both active for channel ops
Open issues / TODOs
- P0 — Update CLAUDE.md §Memory & Tools to reflect 21 MCPs (not 8); link to MCP-INVENTORY
- P1 — Fix
postgres-diag: verify POSTGRES_MCP_DIAG_URL is exported in .mcp.json env block - P1 — Fix
hubspotMCP: verify HUBSPOT_PAT export; test withmcp__hubspot__search_crm - P1 — Fix
stripeMCP: verify STRIPE_SECRET_KEY export - P1 — Fix
google-workspace: reload Claude Code window; verify OAuth token persisted - P2 — Replace
slackMCP: find non-deprecated npm package; add SLACK_TEAM_ID - P2 — Fix
figmaMCP: Phase 11.3a per openclaw-fragmentation-fix-2026-05-01 - P2 — Add perplexity to MCP-INVENTORY.md (currently 21, live shows 22 including perplexity)
- P3 — Build 6 internal MCPs from Tier D backlog (start with checkCompliance — highest ROI, 244 LOC)
- P3 — NemoClaw cred-proxy (B1-B3 ratification → deploy to :18901)
Recent activity
- 2026-05-03: system hub created (W2-S4, Wave 2); live probe shows 22 entries (14 connected, 6 failed); CLAUDE.md drift flagged
- 2026-05-03: google-workspace MCP wired (henry@reri.co OAuth) + airtable MCP confirmed working
- 2026-05-03: MCP-INVENTORY.md v1 created (21 MCPs documented) as part of vendor audit
- 2026-04-21: original 8 MCPs documented in CLAUDE.md §Memory & Tools