mcp-local-rag

Paste it into your assistant. It asks for the manifest, the server wiring and the skills, and for an honest account of what this record does not say. Read that second file first.

Draft an Agent Plugin (agent-plugins.org, specification 1.1.0) that wraps the
MCP server described below, so that somebody could install one thing and have
an assistant that knows when and how to use it.

An Agent Plugin is one installable unit: a `plugin.json` manifest, an
`mcp.json` that wires up the servers it needs, and a `skills/` directory
where each skill is a folder holding a `SKILL.md`. Hand back every file in
full, each under its own path, ready to save.

1. Write `plugin.json` with `$schema` exactly `https://agent-plugins.org/schemas/1.1.0/plugin.schema.json`. The name is
   1 to 64 characters of a-z, 0-9, `-` and `.`, alphanumeric at both ends, with
   no `--` and no `..` in it.

2. Write `mcp.json` wiring THIS server exactly as its record declares it. A
   remote keeps the URL and the transport type as written. A package keeps the
   registry, the identifier and the version as written. Do not invent a command,
   a port, a flag or an argument that is not in the record.

3. Every secret stays an input. No key, token, password or connection string
   belongs in either file. Declare what has to be supplied, name it, and say what
   it is for.

4. Do not invent tools. The record lists the tools it lists, and if it lists
   none then the honest plugin says the tool list was not published rather than
   guessing one from the description.

5. Skills are jobs, not tools. Write one skill per thing somebody would actually
   ask for, and inside each one say when to reach for this server, what a good
   result looks like, and what to do when it comes back empty. A skill per tool
   is a manual page with a different filename.

6. Say where it stops. Name what this plugin will not do — what it has no tool
   for, what needs a person, and what it must not be pointed at. A plugin with no
   stated edge reads as one with no edge.

7. Keep the author's own words for the description. If you would rather say it
   differently, say yours somewhere else and leave theirs where it is.

8. Record which version of the server you wrapped, and where the record came
   from, at the top of `plugin.json`'s description or in the readme. A plugin
   nobody can trace back to a version is one nobody can update.

Produce a second file alongside them, `LIMITS.md`, and treat it as the more
important of the two. The plugin is for whoever installs it. This is for
whoever has to decide whether installing it is a good idea, and that is
usually a different person who will never read the manifest.

It has three parts.

**What this is built from.** One paragraph: whose server it is, what the
record says it does, which version, and the fact that the record is all you
had. Say plainly that nobody ran it.

**What the record does not say.** One entry per gap. Whether the tool list was
published. What the server does with what it reads. What it costs. Whether it
writes anything anywhere. What credentials it will ask for and what those
credentials can reach. An unanswered question stays an unanswered question:
do not fill one in from the description or from what similar servers usually do.

**What a person has to check before trusting it.** The specific things
somebody should verify for themselves, in the order that would stop them
soonest if the answer is bad.

Write it in plain English, and do not soften it. A plugin drafted from a
directory record is a starting point to argue with, not something to install
into anything that matters.

The server record follows, exactly as the public index holds it. It is
everything I have: nobody has run this server, called a tool on it, or checked
that the address answers.

```json
{
  "server": {
    "$schema": "https://static.modelcontextprotocol.io/schemas/2025-12-11/server.schema.json",
    "name": "io.github.shinpr/mcp-local-rag",
    "description": "Easy-to-setup local RAG server with minimal configuration",
    "repository": {
      "url": "https://github.com/shinpr/mcp-local-rag",
      "source": "github"
    },
    "version": "0.18.0",
    "packages": [
      {
        "registryType": "npm",
        "registryBaseUrl": "https://registry.npmjs.org",
        "identifier": "mcp-local-rag",
        "version": "0.18.0",
        "transport": {
          "type": "stdio"
        },
        "environmentVariables": [
          {
            "description": "Base directory for document storage (defaults to current working directory). Ignored when BASE_DIRS is set.",
            "format": "string",
            "name": "BASE_DIR"
          },
          {
            "description": "JSON array of base directories (e.g. '[\"/a\",\"/b\"]'). Takes precedence over BASE_DIR.",
            "format": "string",
            "name": "BASE_DIRS"
          },
          {
            "description": "Path to LanceDB database directory (defaults to ./lancedb/)",
            "format": "string",
            "name": "DB_PATH"
          },
          {
            "description": "Directory where Transformers.js models are cached (defaults to ./models/)",
            "format": "string",
            "name": "CACHE_DIR"
          },
          {
            "description": "Embedding model name (defaults to Xenova/all-MiniLM-L6-v2)",
            "format": "string",
            "name": "MODEL_NAME"
          },
          {
            "description": "Maximum file size in bytes (defaults to 104857600 / 100MB)",
            "format": "string",
            "name": "MAX_FILE_SIZE"
          },
          {
            "description": "Maximum distance threshold for filtering search results. Results with distance greater than this value will be excluded. Lower values mean stricter filtering (e.g., 0.5 for high relevance only)",
            "format": "string",
            "name": "RAG_MAX_DISTANCE"
          },
          {
            "description": "Grouping mode for quality filtering. 'similar' returns only the most similar group (stops at first distance jump). 'related' includes related groups (stops at second distance jump). Unset means no grouping filter",
            "format": "string",
            "name": "RAG_GROUPING"
          },
          {
            "description": "Maximum number of files to keep in search results. Results are filtered to include only chunks from the top N best-scoring files. For example, 1 returns only the single best-matching file's chunks. Unset means no file filtering.",
            "format": "string",
            "name": "RAG_MAX_FILES"
          },
          {
            "description": "Minimum chunk length in characters (1-10000, defaults to 50). Chunks shorter than this threshold are filtered out during ingestion.",
            "format": "string",
            "name": "CHUNK_MIN_LENGTH"
          },
          {
            "description": "Store supported PDF and DOCX images during ingestion and return them with matched chunks (defaults to false).",
            "format": "string",
            "name": "STORE_IMAGES"
          },
          {
            "description": "Execution device for the embedder (defaults to cpu). Passed straight to ONNX Runtime; see the Transformers.js device source for the supported backend names. If the requested device fails to initialize, the server throws an error.",
            "format": "string",
            "name": "RAG_DEVICE"
          },
          {
            "description": "Embedding quantization dtype for the embedder (defaults to fp32). Opt-in and pass-through; accepts any dtype the chosen model provides (fp32, fp16, q8, int8, ...). If the model has no variant for the requested dtype, the server throws an error. Changing this changes the embedding space — re-ingest existing data.",
            "format": "string",
            "name": "RAG_DTYPE"
          },
          {
            "description": "Keyword boost factor for hybrid search (0.0-1.0, defaults to 0.6). 0 means semantic similarity only; higher values increase the keyword-match contribution to the final score.",
            "format": "string",
            "name": "RAG_HYBRID_WEIGHT"
          }
        ]
      }
    ]
  },
  "_meta": {
    "io.modelcontextprotocol.registry/official": {
      "status": "active",
      "statusChangedAt": "2026-08-23T04:33:46.02302Z",
      "publishedAt": "2026-08-23T04:33:46.02302Z",
      "updatedAt": "2026-08-23T04:33:46.02302Z",
      "isLatest": true
    }
  }
}
```

The server is listed as "mcp-local-rag".