mata-kadalz 🦎

Lizard Eyes — local vision MCP server backed by Qwen3-VL-4B running on llama-server. Gives any MCP client (opencode, Claude, Codex, ...) one tool: vision.inspect(image_path, task).

client -> vision.inspect -> mata-kadalz server.py (stdio or streamable HTTP)
                            -> HTTP POST http://<llama-server>:9931/v1/chat/completions
                            -> llama-server (Qwen3-VL-4B GGUF + mmproj)

mata-kadalz is a thin MCP layer. The inference runtime (llama-server from llama.cpp) and the model are external dependencies you install yourself; this repo never downloads or bundles them.

Quick start

Install from source. This project is not yet published to PyPI, so the package is installed from this repository (see Setup guides for your platform).

1. Install Python >= 3.11.

2. Install llama-server (llama.cpp) and the Qwen3-VL-4B model + mmproj — external, from official sources (links in the Model and Setup guides sections).

3. Start llama-server and confirm it is healthy: curl http://localhost:9931/health → {"status":"ok"}.

4. Install mata-kadalz from this repo into a venv:

   git clone https://github.com/kadalzbaiq/mata-kadalz.git
   cd mata-kadalz
   bash scripts/install.sh          # POSIX (Linux/macOS/WSL); Windows: see docs/SETUP-windows.md

5. Confirm the MCP server can reach llama-server: .venv/bin/mata-kadalz --health.

6. Register mata-kadalz in your MCP client — see Client registration.

Related MCP server: Vision MCP Server

Supported setups

Model

• Model: Qwen3VL-4B-Instruct-Q4_K_M.gguf (LLM, ~2.5 GB)

• Vision encoder: mmproj-Qwen3VL-4B-Instruct-F16.gguf (~800 MB)

• Source: Qwen/Qwen3-VL-4B-Instruct-GGUF

  • https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/resolve/main/Qwen3VL-4B-Instruct-Q4_K_M.gguf

  • https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/resolve/main/mmproj-Qwen3VL-4B-Instruct-F16.gguf

Do not change the quant or mmproj combo without re-validating; this is the verified working setup.

Layout

server.py                 # MCP server (stdio + streamable HTTP), single file
config/config.json        # optional overrides (empty = auto)
scripts/install.sh        # install the mata-kadalz package only
docs/                     # per-host setup guides
tests/                    # pytest (no llama-server needed)
runtime/                  # logs + image cache (gitignored)

Client registration

Replace /path/to/mata-kadalz/.venv/bin/mata-kadalz with the real path on your machine.

stdio (local — same machine as the images):

# claude
claude mcp add vision -- /path/to/mata-kadalz/.venv/bin/mata-kadalz
# codex
codex mcp add vision -- /path/to/mata-kadalz/.venv/bin/mata-kadalz

streamable HTTP (remote — server machine may differ from client):

# claude
claude mcp add --transport http vision http://127.0.0.1:9932/mcp
# codex
codex mcp add vision --url http://127.0.0.1:9932/mcp

opencode (config in opencode.json):

{
  "mcp": {
    "vision": {
      "type": "local",                 // stdio
      "command": ["/path/to/mata-kadalz/.venv/bin/mata-kadalz"]
    }
    // or "type": "remote", "url": "http://127.0.0.1:9932/mcp"  // streamable HTTP
  }
}

Usage

One tool: vision.inspect — takes image_path (absolute path on the machine where the MCP server runs) and task (what to analyze). Returns structured JSON:

{ "success": true, "summary": "...", "details": "", "warnings": [], "cache_hit": true }

On failure it returns is_error: true with a machine-readable code, e.g. IMAGE_NOT_FOUND, IMAGE_NOT_SUPPORTED, IMAGE_PATH_NOT_ALLOWED, LLAMA_SERVER_URL_NOT_SET, LLAMA_SERVER_TIMEOUT, LLAMA_BUSY, INVALID_VISION_RESPONSE.

The server also embeds a system prompt (SYSTEM_PROMPT) instructing the vision model how to structure its output; the prompt is sent on every inference request.

HTTP vs stdio — where files must live

• stdio (local): the MCP client and the server share one machine, so image_path is a path on that machine.

• streamable HTTP (remote): the client and the server may be on different machines — image_path is resolved on the server machine, not the client's. Set VISION_IMAGE_ROOTS to restrict which directories the server will read from (strongly recommended for a network-exposed server).

Security for HTTP deployments

If you expose the server over the network:

• Set VISION_IMAGE_ROOTS so the server can only read from the directories you choose. With it unset, the server can read any path on the host.

• Bind to a safe interface. --host 127.0.0.1 (the default) only accepts local connections. For LAN/remote access, prefer a VPN or a firewall rule over binding 0.0.0.0 on a public interface.

• No authentication is built in. Put the endpoint behind an authenticated reverse proxy or your VPN. The HTTP transport speaks raw MCP; there is no token/user layer.

• Allow inbound traffic only on the ports you use: 9931 (llama-server) and 9932 (mata-kadalz HTTP transport).

Configuration

Config precedence: DEFAULTS < config/config.json (empty values skipped) < environment variables.

Example override in config/config.json:

{ "LLAMA_SERVER_URL": "http://192.168.64.1:9931" }

Restrict a network-exposed server to one directory:

{ "VISION_IMAGE_ROOTS": ["/srv/shared-images"] }

Health check

Confirm the MCP can reach llama-server before wiring up a client:

.venv/bin/mata-kadalz --health

Prints platform, resolved LLAMA_SERVER_URL, and a reachable: true/false health probe; exits 0 when reachable.

Caching

Requests are deduplicated by sha256(image) + task + model. A cache hit returns instantly without touching llama-server. Failed requests are never cached. Inference is serialized with a process-wide lock (one concurrent call at a time); the queue beyond the lock is bounded by VISION_MAX_QUEUE and returns LLAMA_BUSY when full.

Cache invalidation: changing VISION_MODEL_ID invalidates the cache automatically, because the model id is part of the cache key — stale answers from an older model are never served.

Cancellation: if the MCP client cancels a request mid-inference, the server releases its lock and queue slot immediately, discards the in-flight result (never caches a partial one), and the error propagates without crashing the server. The underlying llama-server call keeps running in the background; its result is ignored.

Self-check

echo '{"image_path":"/path/to/img.png","task":"describe"}' | .venv/bin/mata-kadalz --once

Reads one JSON request from stdin, runs it, prints the result, and exits: 0 on success or cache hit, 1 on any error (invalid input, missing file, unreachable llama-server, ...). Useful for scripting and cron-style smoke checks.

Test

.venv/bin/python -m pytest

No llama-server required — tests cover config, file validation, magic bytes, cache logic, image-root policy, cancellation, --once exit codes, WSL gateway detection, bounded queue, and a real streamable-HTTP session over uvicorn.

HTTP transport dependency

uvicorn is only needed for --transport http. It already ships transitively with the MCP SDK, but it is also declared as an explicit optional extra so the intent is unambiguous:

# from a source checkout (until PyPI publication)
pip install -e '.[http]'          # or: bash scripts/install.sh then pip install -e '.[http]'

Until the package is published to PyPI, pip install mata-kadalz and pip install 'mata-kadalz[http]' will not work. Use a source checkout.

License

MIT