The author draws the boundary explicitly: this is not a generic SQL server, not a replacement for LangSmith or LangGraph Studio, and not the Agent Server API. It exists for the debugging case where the prompt is fine and the checkpoint is not — and the credentials stay in a profile rather than in a model-facing argument.
A read-only MCP server for inspecting LangGraph checkpoints, thread state and long-term memory, using LangGraph's own checkpointer and store APIs rather than raw SQL.
- Threads and checkpoints inspected — the checkpoint an agent resumed from, the user ID in configurable state, an oversized message history
- Long-term memory read from a PostgresStore, including the namespace a memory was written under
- PostgreSQL, SQLite and Redis checkpointers all supported
- Read-only enforcement in v0.1, so an inspection session cannot mutate state
- Named profiles with environment-variable expansion, keeping database credentials out of the arguments the model sees
- Secret redaction in health checks and error output
- MCP resources with stable readable state URIs, and prompts for repeatable debugging workflows
- Pagination, for threads with more history than one response should carry
A LangGraph persistence layer it can reach — PostgreSQL, SQLite or Redis — configured as a named profile rather than a DSN passed at call time. It runs as a local stdio server for development. Published on PyPI as langmcp.
One command — uvx "langmcp[all]" serve --config ABSOLUTE_PATH_TO_LANGMCP_TOML