MCP server: a claude-context replica

MFS ships a replica of claude-context ("make the codebase the context for any coding agent"), rebuilt on top of MFS and exposed over the Model Context Protocol for any MCP client — Claude Code, Cursor, Codex, Windsurf. Because the index is MFS, one server covers every source you've indexed — code, docs, issues, chat, databases — not just one codebase.

The runnable server is in the example; MFS's search + read API is enough to reproduce it in about 60 lines over the Python SDK.

Two tools

from mcp.server.fastmcp import FastMCP
import mfs_sdk

mcp = FastMCP("claude-context")


@mcp.tool()
def search(query: str, scope: str = "", top_k: int = 8) -> str:
    """Hybrid search across MFS-indexed sources. Empty scope = everything;
    or pass a path / URI prefix like "github://org/repo"."""
    resp = retrieval.search(q=query, path=scope or None, top_k=top_k)
    return "\n\n".join(f"## {h.source}\n{h.content.strip()}" for h in resp.results)


@mcp.tool()
def read(source: str, lines: str = "") -> str:
    """Read a hit in full, or a line range like "40:80", by its source URI."""
    return browse.cat(source, range=lines or None).content


if __name__ == "__main__":
    mcp.run()

search locates, read pulls the exact unit into context — the same loop MFS is built around. (retrieval / browse are mfs_sdk.RetrievalApi / BrowseApi pointed at your server; see the example for the few lines of setup.)

Configure MFS (matches claude-context)

claude-context ships with OpenAI text-embedding-3-small embeddings and a Zilliz Cloud vector database; this example defaults to the same. MFS reads the vector store from the environment and the embedding model from its config, so launch the MFS server with:

export OPENAI_API_KEY=sk-...                              # embeddings
export ZILLIZ_URI=https://<your-cluster>.zillizcloud.com  # vector store
export ZILLIZ_TOKEN=<your-zilliz-key>

# server.toml  (or run: mfs-server setup --section embedding)
[embedding]
provider = "openai"
model    = "text-embedding-3-small"
dim      = 1536

The MCP server is a thin client — it only needs MFS_URL / MFS_TOKEN to reach that server; the embedding and vector-store choice lives with MFS. (Drop these and MFS runs fully local — ONNX embeddings + Milvus Lite, no keys.)

Register it

With Claude Code, from your project:

claude mcp add claude-context \
  --env MFS_URL=http://127.0.0.1:13619 \
  -- uv run --with mcp --with /abs/path/to/mfs/sdks/python python /abs/path/to/mfs/examples/claude-context/server.py

claude mcp list should report claude-context: ✔ Connected, after which the agent calls the tools on its own — "where is rate limiting implemented? search our sources." Any MCP client works; point its stdio server config at the same command.

Restrict its reach

By default the server can search and read everything the MFS server has indexed. Add --env MFS_ALLOWED_SCOPES=github://org/repo,file://local/abs/path (a comma-separated list of URI / path prefixes) to bound it: search only returns hits under those prefixes (an empty scope searches all of them, not the whole index) and read refuses any source outside them. This is enforced by the server, unlike the per-query scope argument the agent chooses.