lean-mcp

LeanMCP is a local MCP chat runtime that turns tool selection into retrieval using a SQLite-backed vector index.

No Bedrock dependency is required for the retrieval layer.

End-to-End Flow (one user message)

0) System boot (once)

1. Start MCP server (stdio).
2. Connect through MCP client.
3. Call tools/list.
4. Convert MCP tools into internal tool objects.
5. Index tools in SQLite vector store.

Result: you have a searchable tool index.

1) User sends message

Example:

show me files in this directory

2) Retrieve relevant tools

Run vector retrieval over tool metadata in SQLite and get top-K relevant tools.

Example output:

[list_directory, read_file]

3) Build model/tool context

Send chat state + user input + only retrieved tools to the decision step.

4) Decide action

Either:

• answer directly, or
• emit tool call (name + args)

5) Execute tool via MCP

Forward the chosen tool call to MCP tools/call.

6) Feed tool result back

Append tool output back into the working context.

7) Generate final answer

Return user-facing response.

8) Multi-step loop (if needed)

Repeat retrieval -> decision -> tool execution until completion or max rounds.

What is different from traditional MCP

• Traditional: LLM sees all tools every turn.
• LeanMCP: runtime injects only top-K retrieved tools every turn.

Everything else (tool execution loop via MCP) stays standard.

Download and run

1) Clone

git clone https://github.com/diyagk01/lean-mcp
cd lean-mcp

2) Install

pip install -e .

Optional local ONNX embeddings:

pip install -e ".[fastembed]"

3) Start chat CLI

You need an MCP server command:

lean-mcp-chat --mcp-command python --mcp-args "path/to/your_mcp_server.py --stdio" --db-path rag_mcp.db

Run one query and exit:

lean-mcp-chat --mcp-command python --mcp-args "path/to/your_mcp_server.py --stdio" --query "show me files in this directory"

CLI commands

• q, exit: quit
• help: show help
• sync: re-list tools from MCP and refresh SQLite index
• tools: list indexed tools
• history: show conversation history
• clear: clear conversation history

Core modules

• rag_mcp/mcp_stdio.py: stdio MCP client (initialize, tools/list, tools/call)
• rag_mcp/runtime.py: boot + per-message retrieval/decision/tool loop
• rag_mcp/chat_cli.py: interactive chat interface