Add LiteRT-LM MCP Server

.github/workflows/verify-litert-mcp.yml

@@ -0,0 +1,35 @@
+name: Verify LiteRT-LM MCP
+
+on:
+  push:
+    paths:
+      - 'mcp-servers/litert-mcp/**'
+      - '.github/workflows/verify-litert-mcp.yml'
+  pull_request:
+    paths:
+      - 'mcp-servers/litert-mcp/**'
+      - '.github/workflows/verify-litert-mcp.yml'
+
+jobs:
+  verify:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+
+      - name: Verify Server Startup (Initialize)
+        run: |
+          echo '{"jsonrpc": "2.0", "id": 1, "method": "initialize", "params": {}}' | python3 mcp-servers/litert-mcp/server.py
+
+      - name: Verify Tools List
+        run: |
+          OUTPUT=$(echo '{"jsonrpc": "2.0", "id": 2, "method": "tools/list", "params": {}}' | python3 mcp-servers/litert-mcp/server.py)
+          echo "$OUTPUT"
+          if [[ "$OUTPUT" != *"run_inference"* ]]; then
+            echo "Error: run_inference tool not found in output"
+            exit 1
+          fi

AGENTS.md

@@ -0,0 +1,23 @@
+# AGENTS.md
+
+## Mission
+To enable autonomous AI agents to efficiently reason, plan, and execute tasks within this repository, specifically focusing on the integration and orchestration of Model Context Protocol (MCP) servers and Video Intelligence pipelines.
+
+## Scope of "Actionable Room"
+Agents are granted permission and encouraged to:
+1.  **Modify and Expand `mcp-servers/`**: Create new MCP servers, update existing ones, and refactor code to improve modularity and performance.
+2.  **Improve Automation**: Create and edit GitHub Actions workflows (`.github/workflows/`) to add robust testing and verification for new features.
+3.  **Refactor for Clarity**: Improve documentation (READMEs) and code structure to facilitate better "Machine Readability" and "Human Understandability".
+
+## Protocols
+1.  **Verify Before Submit**:
+    - Always run relevant verification scripts or tests before submitting changes.
+    - If no test exists for a new feature, **create one**.
+2.  **CI/CD Alignment**:
+    - Ensure all changes pass existing CI checks.
+    - When adding a new component (like an MCP server), add a corresponding CI workflow to ensure it remains functional.
+3.  **Cross-Platform Compatibility**:
+    - Write code that is compatible with Linux and Windows environments whenever possible (e.g., handling `asyncio` loops correctly).
+4.  **Documentation**:
+    - Update `README.md` files when interface changes occur.
+    - Document limitations (e.g., "Text-only CLI wrapper") clearly.

mcp-servers/litert-mcp/README.md

@@ -0,0 +1,61 @@
+# LiteRT-LM MCP Server
+
+This MCP server provides an interface to Google's **LiteRT-LM**, a high-performance runtime for Large Language Models (LLMs) on edge devices (Android, iOS, Linux, MacOS, Windows).
+
+It allows you to run inference on local models (like Gemma, Phi, Qwen) directly from your MCP ecosystem.
+
+**Note:** This server currently wraps the `lit` CLI. Multimodal inputs (image/audio) are enabled in the interface but require the C++ API or Python bindings. The CLI wrapper currently supports **Text-only inference** until CLI flags for multimodal are verified.
+
+## Prerequisites
+
+1.  **LiteRT-LM**: You must have LiteRT-LM installed or built.
+    *   Official Repository: [google-ai-edge/LiteRT-LM](https://github.com/google-ai-edge/LiteRT-LM)
+    *   Follow the "Build and Run" instructions in the official repo to build the `lit` CLI or `litert_lm_main` binary.
+    *   Alternatively, download prebuilt binaries if available for your platform.
+
+2.  **Models**: Download a supported `.litertlm` model.
+    *   Models are available on Hugging Face: [LiteRT Community](https://huggingface.co/litert-community)
+
+## Configuration
+
+Set the following environment variables:
+
+*   `LIT_BINARY_PATH`: Path to the `lit` CLI executable or `litert_lm_main` binary. Defaults to `lit` (assuming it's in your PATH).
+*   `LIT_MODEL_PATH`: Default path to your `.litertlm` model file. (Optional, can be passed per request).
+
+## Usage
+
+### Tools
+
+#### `run_inference`
+
+Runs inference using the configured LiteRT-LM model.
+
+*   **Arguments**:
+    *   `prompt` (string, required): The input text prompt.
+    *   `model_path` (string, optional): Path to the `.litertlm` model file. Overrides `LIT_MODEL_PATH` env var.
+    *   `image_path` (string, optional): Path to an image file for multimodal inference.
+    *   `audio_path` (string, optional): Path to an audio file for multimodal inference.
+    *   `backend` (string, optional): Backend to use (`cpu`, `gpu`, `npu`). Defaults to `cpu`.
+
+### Example
+
+```json
+{
+  "name": "run_inference",
+  "arguments": {
+    "prompt": "What is the capital of France?",
+    "backend": "cpu"
+  }
+}
+```
+
+## Setup for Development
+
+This server uses a manual JSON-RPC implementation to avoid external dependencies in the base environment. Just run:
+
+```bash
+python3 server.py
+```
+
+Ensure `LIT_BINARY_PATH` is set correctly.