MCP Tool Catalog for Boss Agent Services
This catalog summarizes the Model Context Protocol (MCP) tools generated from the Boss Agent unified registry. Each tool uses registry metadata (friendly names, Cloud Run base URLs, and default process endpoints) so callers can invoke services without hard-coded paths.
Registry-backed service tools
| Tool name | Service | Default endpoint | Description | Example payload |
|---|---|---|---|---|
invoke_graph_writer |
Graph Writer | /process |
RETIRED AT INVOKE (2026-08-22, owner-approved). The backend (Neo4j) was retired 2026-08-09; with the mounts stripped, writer.py dials bolt://localhost:7687 and every write lands nowhere while the app answers "accepted" — a phantom write. Both boss transports now refuse the service id before any HTTP (_RETIRED_BACKEND_SERVICES, pinned by tests/test_kg_write_perimeter.py); the registration stays so tool counts hold. Journey/profile writes ride the governed ingestion door; KG mutations use the authenticated /api/v1/kg/* routes. Lifting the refusal requires a governed repoint, never a bare revert (design: docs/architecture/GOVERNED_REPOINT_RETIRED_TOOLS.md). |
{ "payload": { "customer_id": "C123", "events": ["signup"] } } |
invoke_predictions_api |
Predictions & Risk Scoring | /process |
RETIRED AT INVOKE (2026-08-24, owner-approved). Measured from service.py: every read path is a Cypher query with no non-Neo4j fallback, and the backend was retired 2026-08-09 with the mounts stripped live 2026-08-10 — the tool can never answer. Both boss transports refuse the service id (_RETIRED_BACKEND_SERVICES, pinned by tests/test_kg_write_perimeter.py); the registration stays so tool counts hold. Journey-level risk scoring does not exist; the governed risk input is get_risk_signal (customer-level churn risk via customer_intelligence_api, estimand labeled — implemented 2026-08-25 per docs/architecture/GOVERNED_REPOINT_RETIRED_TOOLS.md); flows must not substitute ungated actions for the missing journey-level score. |
{ "payload": { "query": "risk for cart abandon" } } |
invoke_lift_engine |
Lift Engine | /api/v1/uplift |
Causal uplift scoring and heterogeneous treatment effects. | { "payload": { "query": "estimate uplift for Q1 campaign" } } |
invoke_ekis |
EKIS | /api/v1/query |
Event-Knowledge Integration System search and resolution. | { "payload": { "query": "lookup customer 123" } } |
invoke_websocket_gateway |
WebSocket Gateway | /api/v1/broadcast |
Real-time broadcast and push messaging. | { "payload": { "channel": "alerts", "message": "deploy complete" } } |
invoke_internal_email_agent |
Internal Email Agent | /api/v1/send |
Automated email composition and delivery. | { "payload": { "to": "ops@example.com", "subject": "Release", "body": "Staging deploy finished" } } |
invoke_gemini_kg_pipeline |
Gemini KG Pipeline | /api/v1/extract |
Gemini-powered KG entity and relation extraction. | { "payload": { "text": "Nike acquired XYZ", "schema": {"entities": ["ORG"]} } } |
invoke_graphrag_integrated |
GraphRAG Integrated | /api/v1/query |
Graph-aware retrieval augmented generation. | { "payload": { "query": "latest marketing insights" } } |
invoke_customer_journey_system |
Customer Journey System | /api/v1/journey |
Journey tracking and touchpoint analytics. | { "payload": { "customer_id": "C123", "action": "track" } } |
invoke_mizoki_journey_kg |
MizOki Journey KG | /api/v1/paths |
RETIRED AT INVOKE (2026-08-25, owner-approved). Measured: the registered Cloud Run service does not exist — the live-generated docs/service_inventory.md lists its siblings graph-writer and predictions-api but has no mizoki-journey-kg row, no source or deploy path ever existed in-repo, and the registered /api/v1/paths endpoint is implemented nowhere in the tree. Both boss transports refuse the service id (_RETIRED_BACKEND_SERVICES, pinned by tests/test_kg_write_perimeter.py); the registration stays so tool counts hold. Journey path analysis: invoke_eshkg_path_metrics / invoke_path_metrics_bq. Repoint design: docs/architecture/GOVERNED_REPOINT_RETIRED_TOOLS.md. |
{ "payload": { "start_node": "user:C123" } } |
invoke_eshkg_path_metrics_api |
E-SHKG Path Metrics API | /api/v1/metrics |
Journey and attribution metrics. | { "payload": { "path_type": "conversion" } } |
invoke_path_metrics_api_bq |
Path Metrics API (BigQuery) | /api/v1/query |
BigQuery-backed path analytics and SQL metrics. | { "payload": { "sql": "SELECT * FROM metrics LIMIT 10" } } |
invoke_unified_revenue_ingestion |
Unified Revenue Ingestion | /api/v1/ingest |
Shopify/Stripe/GA4 revenue ingestion. | { "payload": { "source": "shopify", "data": {"order_id": "1"} } } |
invoke_vertex_ai_integration |
Vertex AI Integration | /api/v1/predict |
Vertex AI / Gemini model serving. | { "payload": { "model": "text-bison", "instances": ["hi"] } } |
invoke_customer_intelligence_api |
Customer Intelligence API | /api/v1/customer |
Customer intelligence, CLV, churn, and segmentation. | { "payload": { "customer_id": "C123" } } |
Tip: The generic
invoke_external_servicetool remains available for ad-hoc calls using explicitservice_id,endpoint, andpayloadparameters.
Routing behavior
- Base URLs and default endpoints are sourced from the unified registry for each service, avoiding hand-coded paths.
- Endpoints are normalized (leading
/enforced) and can be overridden per call if needed. - MCP client definitions (
mcp/src/server.ts) mirror this catalog so external tools are discoverable vialist_tools.
Boss Agent ADK source of truth
Boss Agent ADK's MCP server is the authoritative source for tool discovery and routing metadata. Use list_tools from the Boss Agent ADK MCP server to validate the active catalog before invoking tools.