zivshek/rag-mcp: a one-file MCP server whose domain lives in its tool description

The whole server is one script at the root

There is no package structure here. The repository holds a single `server.py` beside `pyproject.toml`, a `uv.lock`, a `.python-version` file, the README and the license, with no source directory and no test directory. That is the whole surface area, and it shapes how you run it: the MCP configuration refers to a path to the server script rather than to a console entry point, so the script is executed directly. The presence of a lock file and of a pinned Python version file says the intended toolchain is uv rather than pip, which matches the run command in the editor configuration, where the script is launched through `uv run`.

The subject matter is chosen by editing prose

The README is blunt about what this tool is and what it is not. It retrieves relevant information about the Godot game engine from a local ChromaDB database, and it says that although the tool is for Godot in particular, with small changes to the tool description it can be used for any Chroma database. That single sentence describes the whole extension mechanism. There is no domain setting, no collection-of-topics mapping and no index of terms, so retargeting the server means editing a piece of text that a model reads. That is a real advantage when you already have a vector store, and it is also why nothing about the tool constrains what may be asked about it.

Ingestion belongs to a different repository

The usage list has three steps, and the first one is not about this project. You get the ChromaDB vector store ready using the web2embeddings tool, a separate repository under the same account, then download or clone this repository to your local disk, and only then configure the MCP server in the editor. The configuration makes the division concrete: the path passed for the database argument is illustrated as a `vector_stores\chroma_db` directory inside that other tool's artifacts. So this repository never creates or updates embeddings. It reads a store that already exists, and if you change the source material you have to go back to the tool that built it, rebuild the store, and only then will this server see different answers.

The documented configuration is written for Windows

The settings block is given as a fragment to paste into the editor's settings.json, under an `mcp` key with an empty `inputs` array and a `servers` object. The server entry is named `godot_rag`, with a comment saying you can use whatever name you want, and the arguments carry their own inline instructions.

"mcp": {
    "inputs": [],
    "servers": {
        "godot_rag": { // or whatever name you want
            "command": "cmd", // replace this with "uv" on Mac or Linux
            "args": [
                "/c", // remove this on Mac or Linux
                "uv", // remove this on Mac or Linux
                "run",
                "path to the server script 'server.py'", // C:\\dev\\rag-mcp\\server.py
                "-d",
                "path to the chroma_db on your computer", // C:\\dev\\web2embeddings\\artifacts\\vector_stores\\chroma_db
                "-c",
                "na

Read the comments rather than the values. The command is `cmd`, which exists only on Windows, and the first two arguments exist only to chain through it. On Mac or Linux you replace the command with `uv` and delete both, which leaves `uv run` followed by the script path. The visible configuration also cuts off after a `-c` argument, so the collection or query naming it takes is not shown in what is published here.

Python 3.13 is the floor, and embeddings are local

The project metadata is short and states the real requirements. Python 3.13 or newer is required, which is a high floor to inherit this early. Three dependencies carry the work: `chromadb` at 0.6.3 or newer for the vector store, `mcp` with its command line extra at 1.6.0 or newer for the protocol server itself, and `sentence-transformers` at 4.0.1 or newer. That last one is the reason the tool needs no service credentials, since it does its embedding work locally rather than calling out to a hosted model. The package is version 0.1.0, and the description field still holds the scaffolding placeholder rather than the sentence the README opens with.

VS Code Insiders is called out as a requirement

One prerequisite is stated as a note rather than a dependency: you may need to download Visual Studio Code Insiders to use MCP or Agent mode. That makes sense given where the configuration lives, since the settings block described above is an editor setting rather than a standalone file, and the Insiders channel is named as the one where this works. The shape of the configuration is worth noting on its own. Servers are declared in a map, so more than one can be registered alongside each other under the same key, the empty `inputs` array suggests shared inputs are optional, and the server name is arbitrary, so a second store could be added as a second entry without touching the first.

No releases, and a project description that was never written

The state of the repository is worth recording before you build on it. The last push to main is dated 2025-04-04 and there are no GitHub releases, so the version you get is whatever the main branch holds. The package version in the metadata is 0.1.0, and the same metadata carries the placeholder description text that scaffolding tools generate, which is a small sign of how little time went into packaging compared with the server itself. The repository is licensed under MIT and has a description field of its own left empty. Nothing in the README describes testing, error handling or how the retrieved context is assembled into an answer, so those questions are answered by reading the one file that does the work.

Editorial conclusion

Use this server if you already have a ChromaDB vector store of documentation and want an editor to query it over MCP, since it does exactly one job and the ingestion side is deliberately somebody else's problem. The porting story is better than the name suggests, because the domain is chosen by editing the tool description rather than by configuration, so any collection built the same way can be served the same way. Two limits before you start. The documented settings block is written for Windows and needs edits on Mac or Linux, and the vector store has to be built first with the separate web2embeddings tool. The last push to main is dated 2025-04-04 and no releases are published.

Frequently asked questions

Official sources

1. Official README
2. Project repository
3. README
4. Releases