{"id":"ogham-mcp-ogham-mcp","name":"ogham-mcp","af_score":53.2,"security_score":39.0,"reliability_score":35.0,"what_it_does":"Ogham MCP is a Python-based Model Context Protocol (MCP) server that provides persistent, cross-client “shared memory” for AI coding agents. It stores memories in a Postgres-compatible database (Supabase/Postgres), generates embeddings via configurable providers (OpenAI/Ollama/Mistral/Voyage/Gemini or local ONNX), and exposes MCP tools for memory storage/retrieval plus additional capabilities like search, graph/profiles, and import/export. It can run per-client in stdio mode or as a shared persistent server via SSE transport, and includes CLI utilities (init/health/search/store/list/export/import/serve/openapi).","best_when":"You want agent-accessible persistent memory with semantic + keyword retrieval, using Postgres/Supabase as the source of truth, and you can manage database + embedding provider credentials.","avoid_when":"You need strong multi-tenant security boundaries enforced by the MCP server (beyond database/security posture), or you cannot run/connect to a database. Also avoid exposing the SSE endpoint broadly without network controls.","last_evaluated":"2026-03-30T13:49:48.603308+00:00","has_mcp":true,"has_api":false,"auth_methods":["Environment-variable based secrets for database and embedding providers (e.g., SUPABASE_KEY for Supabase service role, DATABASE_URL for Postgres, provider API keys)","No explicit MCP client authentication/authorization mechanism described in provided README"],"has_free_tier":false,"known_gotchas":["If running in SSE mode, clients point to the shared server URL; ensure you run it persistently and control access to the SSE endpoint.","Embedding provider configuration (dimensionality/EMBEDDING_DIM and schema vector(N)) must match; mismatches can lead to failures when storing/searching.","Temporal query parsing is mostly local (parsedatetime) but can fall back to an LLM if configured; this can change cost/latency.","Secret masking is described for hook-driven inscription, but agents should still avoid sending raw secrets in memory-related tool calls if possible."],"error_quality":0.0}