Headless, runtime-agnostic MCP server serving knowledge retrieval, belief management, skill compilation, and tool audit logging. Provides vector search, RAG synthesis with citations, feedback signaling, and external tool-use audit trails for agent assurance.
AgenticMind defines 13 tools with reasonable naming (verb_noun pattern: kl_search, mem_write, kl_ingest) and descriptions (avg ~120 chars, meeting 10-1024 baseline). However, critical gaps exist: input schemas are inferred from the provided JSON snippets but NOT directly visible in source code (only tool names and descriptions shown in mcp-tools.ts). Parameter descriptions are present for kl_search and kl_ask_global but missing for most others (mem_write, mem_read, mem_retract, kl_ingest, kl_remove, skill_compile, hitl_* tools show no parameter details). Output schemas are undocumented. Error handling and recovery guidance are absent. Security patterns (scope declarations, audit trails) are mentioned in schema but not enforced in tool definitions. Composition is reasonable (single-responsibility tools), but chaining IDs are not guaranteed in responses.
Answer a pending human-in-the-loop request. Provides the human decision/response to an open HITL request.
Retrieve the status and details of a human-in-the-loop request by id.
Create a human-in-the-loop request for manual review or decision. Stores the request with context and awaits human response.
RAG synthesis with citations. Synthesizes answers tailored to the calling agent's goal, accounting for arbitrary caller facts, with optional faithfulness judgment and contested-sources detection.
Query the knowledge graph for neighbors of a given entity. Returns connected entities and their relationships. Only available when a GraphStore is configured.
Ingest text, PDF, DOCX, or XLSX material into the knowledge base. Chunks, embeds, and stores with optional acceptance evaluation and skill extraction. Requires kl_ingest scope and blob store.
Input schemas missing for 11 of 13 tools.
Output schemas completely undocumented. No tool describes what fields it returns, their types, or structure. LLMs cannot plan downstream calls or extract required data (e.g., does kl_search return chunk_id, source, score, text?). Violates pattern:tool and pattern:response-shaper.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 46 | 2026-07-28+ | v2 |
Remove ingested material from the knowledge base by material id. Deletes all chunks and embeddings associated with the material.
Vector retrieval over the chunks index. Accepts a batch of sub-queries (queries) fanned out in one round-trip, and an optional tokenBudget to return the best ~N tokens of context instead of a fixed count.
Feedback signal on a prior kl_ask_global answer. Records agent-supplied correctness/relevance/hallucination signals with optional explanation and source citations for tuning the retrieval/synthesis pipeline.
Recall beliefs from the agent's private memory. Retrieves stored facts with their strength, source citations, and lifecycle metadata.
Retract a belief from the agent's private memory. Removes a stored fact by id.
Write a belief to the agent's private memory. Stores a fact with optional source citations, strength, and lifecycle metadata. Requires mem_write scope.
Compile a skill definition (YAML) into executable code. Validates syntax, extracts dependencies, and generates runnable skill implementations.
No error handling or recovery guidance. Tools lack descriptions of failure modes, retryability, or actionable error messages. E.g., kl_ingest (WRITE) and kl_remove (DESTRUCTIVE) have no confirmation or dry-run patterns. Violates pattern:recovery-guide and pattern:confirmation-request.
Parameter descriptions missing or minimal for 11 tools. mem_write, mem_read, mem_retract, kl_ingest, kl_remove, kl_graph_neighbors, skill_compile, hitl_* tools show only brief descriptions (40-60 chars) with no format, range, or constraint guidance. Violates pattern:tool-description baseline (avg 72 chars per param).
No scope declarations or permission gates visible. Tools like kl_ingest (WRITE), kl_remove (DESTRUCTIVE), mem_write (WRITE) lack documented permission requirements (e.g., 'kl_ingest scope', 'write:memory'). Audit trail schema exists (tool_audit_events table) but not enforced in tool definitions. Violates pattern:scope-declaration and pattern:permission-gate.