HTTP MCP server exposing the pgvector learnings DB. Replaces direct psql/ssh access used by rag-query, learn-seed, learn-inject, learn-context.

Transport: streamable HTTP (/mcp)
Port: 9221
Backend: learnings_embeddings on the infra postgres container
Embeddings: LiteLLM /v1/embeddings (text-embedding-3-small, 1536 dims)

Tools

Tool	Replaces
`learning_query`	`rag-query`
`learning_search_fts`	`pg_search_fts` in `pglearn.sh`
`learning_inject`	`learn-inject`
`learning_context`	`learn-context`
`learning_add`	`pg_insert_learning`
`learning_mark_applied`	`pg_mark_applied`

Env

PGHOST=postgres        # docker service name on infra
PGPORT=5432
PGUSER=agiliton
PGPASSWORD=...         # from vault ag.postgres.agiliton_password
PGDATABASE=agiliton
LITELLM_URL=https://llm.agiliton.cloud
LITELLM_API_KEY=...    # vault litellm.ai_coder_key or similar
EMBEDDING_MODEL=text-embedding-3-small
SENTRY_DSN=            # optional

Local dev

npm install
LITELLM_API_KEY=$(vault get litellm.ai_coder_key) \
PGHOST=localhost PGPORT=5433 \
node src/http-server.js

curl -s http://127.0.0.1:9221/health

Deploy

Built as gitea.agiliton.internal:3000/christian/learning-mcp:latest and attached to the infra docker-compose stack with the postgres network. Registered in /opt/apps/litellm/config.yaml as a streamable-HTTP MCP server (CF-3081 pattern).