LangMCP

Read-only MCP server for inspecting LangGraph checkpoints, thread state, and long-term memory.

LangMCP helps you answer the debugging question that traces do not always answer:

What is actually saved in my LangGraph persistence layer right now?

It is not a generic SQL MCP. It uses LangGraph-native checkpointer and store APIs, connects through named profiles, and keeps database credentials out of tool arguments.

Why LangMCP

When a stateful agent behaves strangely, the problem is often not only the prompt. It might be the checkpoint it resumed from, the user ID in configurable state, the store namespace used for memory, or an oversized message history.

LangMCP gives MCP clients such as Cursor and Claude Desktop a safe inspection surface for those questions.

Use LangSmith for traces, LangGraph Studio for visual graph workflows, and LangMCP when you want an assistant in your editor to inspect persisted state through a narrow read-only interface.

Features

• Profile-based config with environment variable expansion.
• Read-only enforcement in v0.1.
• Secret redaction in health checks and error output.
• PostgreSQL, SQLite, and Redis checkpointer inspection.
• PostgreSQL PostgresStore long-term memory inspection.
• MCP tools for threads, checkpoints, store data, and analysis.
• MCP resources for stable readable state URIs.
• MCP prompts for repeatable debugging workflows.
• Pagination and truncation for large responses.

Installation

uv pip install "langmcp[all]"

Or run without installing:

uvx "langmcp[all]" --version

LangMCP supports Python 3.11 and 3.12. The repository includes a .python-version file set to Python 3.12.

Configuration

Copy the example config and environment files:

cp examples/langmcp.example.toml langmcp.toml
cp .env.example .env

Set a read-only database URI in .env:

POSTGRES_URI=postgresql://READONLY_USER:READONLY_PASSWORD@HOST:5432/DB_NAME
LANGMCP_READ_ONLY=true

LangMCP loads .env automatically when present. Existing shell environment variables take precedence.

Example langmcp.toml:

[defaults]
profile = "dev"
read_only = true
max_response_chars = 250000

[profiles.dev]
checkpointer = "${POSTGRES_URI}"
store = "${POSTGRES_URI}"

[profiles.local_sqlite]
checkpointer = "sqlite:///./.langgraph/checkpoints.db"

[profiles.local_redis]
checkpointer = "redis://localhost:6379/0"

Environment overrides:

• LANGMCP_CONFIG
• LANGMCP_PROFILE
• LANGMCP_READ_ONLY
• POSTGRES_URI
• LANGMCP_CHECKPOINTER_URI
• LANGMCP_STORE_URI

Verify Setup

Run:

langmcp doctor --config ./langmcp.toml

The doctor command checks connectivity, backend types, setup status, package versions, and redacts sensitive URI fields.

Cursor Setup

See examples/cursor-mcp.json.

Minimal shape:

{
  "mcpServers": {
    "langmcp": {
      "command": "uvx",
      "args": ["langmcp[all]", "serve", "--config", "ABSOLUTE_PATH_TO_LANGMCP_TOML"],
      "env": {
        "LANGMCP_READ_ONLY": "true",
        "POSTGRES_URI": "postgresql://READONLY_USER:READONLY_PASSWORD@HOST:5432/DB_NAME"
      }
    }
  }
}

Start the server directly:

langmcp serve --config ./langmcp.toml

Example Assistant Prompts

Once connected through MCP, ask your assistant:

Use LangMCP to summarize thread THREAD_ID and check whether user memory exists for USER_ID.

Compare checkpoint CHECKPOINT_A and CHECKPOINT_B for thread THREAD_ID. Tell me what changed.

Analyze whether thread THREAD_ID is carrying too much context.

Investigate a possible memory gap for thread THREAD_ID and user USER_ID.

MCP Tools

All tools accept optional profile unless noted. Responses include profile, truncated, and pagination fields where applicable.

MCP Resources

Resources expose readable state through stable MCP URIs.

For multi-part namespaces, prefer the get_store_item tool if your MCP client treats / as a path separator inside resource parameters.

MCP Prompts

Prompts package repeatable investigations.

Backend Matrix

Security

1. Tools accept profile names, not raw DSNs.
2. read_only=true is enforced in v0.1.
3. Use a read-only PostgreSQL user for shared environments.
4. Passwords are redacted in health_check and CLI output.
5. Redis thread discovery uses SCAN with limits. Avoid broad scans on very large instances.
6. Commit examples/langmcp.example.toml and .env.example, not real langmcp.toml or .env files.

Development

uv pip install -e ".[all,dev]"
ruff check .
pytest tests/unit -v

Integration tests use local Docker services:

docker compose -f docker-compose.test.yml up -d
POSTGRES_URI=postgresql://langgraph:langgraph@localhost:5442/langgraph \
  REDIS_URI=redis://localhost:6379/0 \
  pytest tests/integration -v -m integration

Use the local test values from docker-compose.test.yml. They are for Docker integration tests only.

Roadmap

• LangGraph Agent Server adapter.
• HTTP transport with team auth.
• Vector store inspection tools.
• Carefully scoped write workflows such as update_thread_state and resume_thread.

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md.

Good first issue ideas:

• Add examples for a specific LangGraph persistence backend.
• Improve error messages for unsupported store backends.
• Add a resource or prompt test for an edge case.

License

MIT. See LICENSE.