tres-gemmas
This Gemini extension exists to make it effortless to run high-performance, on-device machine learning models.
Core trick
Node.js native bindings wrap the LiteRT-LM C++ runtime, exposing on-device model inference on CPU or GPU through an mcp-llm-compatible API.
Tags
Description
This Gemini extension exists to make it effortless to run high-performance, on-device machine learning models. It packages the LiteRT-LM runtime, GPU-first bindings, and MCP-compatible tooling into a single, hackathon-friendly bundle so teams can focus on building experiences rather than wiring up infrastructure.
README excerpt
# tres vibes: accelerated model serving over mcp
Native Node.js bindings for [LiteRT-LM](https://github.com/google/litert-lm) with an API matching the `mcp-llm` Rust crate from [assist-mcp](~/assist-mcp).
## Features
- **Native Performance**: Direct bindings to LiteRT-LM C++ library via node-gyp
- **TypeScript Support**: Full TypeScript types and IDE autocomplete
- **mcp-llm Compatible API**: Same interface as the Rust `mcp-llm` crate
- **Conversation Management**: Multi-turn conversations with context
- **Benchmarking**: Built-in performance metrics (tokens/sec, latency)
- **CPU & GPU**: Support for both CPU and GPU backends
## Prerequisites
1. **Prepare LiteRT-LM runtime artifacts**
This repository expects prebuilt LiteRT-LM headers and shared libraries inside `prebuilt/<platform>/`.
To build them yourself from the LiteRT-LM sources, run:
```bash
cd ~/LiteRT-LM
bazel build //rust_api:litert_lm_rust_api \
--define=litert_link_capi_so=true \
--define=resolve_symbols_in_exec=false
```
Copy the resulting headers and shared libraries into this repo:
```bash
# Headers
mkdir -p /Users/rpm/gemma3-mcp/prebuilt/include
cp rust_api/litert_lm_c_api.h /Users/rpm/gemma3-mcp/prebuilt/include/
# macOS example (adjust for linux/windows)
mkdir -p /Users/rpm/gemma3-mcp/prebuilt/darwin/lib
cp bazel-bin/rust_api/*.dylib /Users/rpm/gemma3-mcp/prebuilt/darwin/lib/
```
The build system will automatically copy these `.dylib/.so/