MCP server for FuseMomo behavioral intelligence API - enables AI agents to resolve identities, log interactions, and get behavioral knowledge
FuseMomo MCP server demonstrates solid tool design with complete schemas, clear descriptions, and proper parameter typing. All 5 tools have explicit Zod schemas with validation. Descriptions are well-written (100-300 chars) and explain WHEN to use each tool. Tool names follow verb_noun convention (resolve_entity, log_interaction, get_recommendation). However, output schemas are not documented in the source code provided, only input schemas are visible. Error handling is present but generic. Tool composition is strong: resolve_entity must be called first (documented), and tools chain logically. Security is good: no secrets in parameters, API key injected server-side via config. Minor gaps: no enum constraints for some string fields, no pagination support for list-like operations, and no explicit error recovery guidance.
Retrieve the complete behavioral profile and consolidated history of an entity. Provides historical success rates, behavioral scores, all linked identifiers, and recent interactions. Use this when you need absolute context on an entity's past behaviors before deciding how to approach them.
Ask the intelligent engine for the statistically best action to take next for a given entity and intent. Based on historical data, this tool returns the exact communication channel or action type most likely to succeed. Call this *before* deciding what to do.
Chronicle a specific action taken on an entity to build its behavioral profile. Use this to explicitly tell the system what you did, which tool you used, and what happened. This behavioral payload forms the graph that drives future ML recommendations.
Resolve external identifiers (e.g. Email, GitHub ID, Stripe ID) into a unified internal Fusemomo Entity. THIS MUST BE YOUR FIRST STEP in any workflow, as every other tool requires the internal `entity_id` returned here. If the entity does not exist, this tool automatically safely creates it.
Output schemas not documented. Tool descriptions explain what is returned (e.g., 'returns the exact communication channel or action type'), but the actual response structure is not visible in source code. LLMs cannot plan downstream tool calls without knowing response fields.
No enum constraints for string fields that should be restricted. 'api' parameter in log_interaction accepts any string (e.g., 'sendgrid', 'stripe', 'linear') but is not declared as an enum. LLMs may hallucinate invalid API names.
Error handling is present but lacks recovery guidance. The client.ts file shows error creation (createErrorFromStatus, ServerError, TimeoutError) but no tool-level error responses guide the LLM on what to do next (retry, ask user, call alternative tool).
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | <=2025-11-25 | v2 |
Close the feedback loop on a previously retrieved recommendation. Once you act upon a recommendation from get_recommendation, you MUST call this to inform the engine if the recommendation actually worked or failed. This self-corrects the algorithm.
No pagination or result limits documented. get_entity and get_recommendation may return large behavioral profiles or recommendation lists, but no limit, offset, or cursor parameters are visible. Large responses risk context window exhaustion.
Tool descriptions lack explicit dependency hints. While resolve_entity is documented as 'FIRST STEP', other tools do not state 'Call resolve_entity first to obtain entity_id' in their descriptions. This forces LLMs to infer the dependency chain.