# open-web-agent-rs
[![Rust](https://github.com/seemueller-io/open-web-agent-rs/actions/workflows/main.yml/badge.svg)](https://github.com/seemueller-io/open-web-agent-rs/actions/workflows/main.yml)
[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](https://opensource.org/licenses/MIT)

A Rust-based web agent with an embedded OpenAI-compatible inference server (supports Gemma models only). It is packaged and deployed as a container.

## Project Structure

This project is organized as a Cargo workspace with the following crates:

- `crates`
  - [agent-server](crates/agent-server): The main web agent server
  - [inference-engine](crates/inference-engine): An embedded OpenAI-compatible inference server for Gemma models
- `packages`
  - [genaiscript](packages/genaiscript): GenaiScript scripts  
  - [genaiscript-rust-shim](packages/genaiscript-rust-shim): An embedded OpenAI-compatible inference server for Gemma models

## Acknowledgements

Special gratitude and thanks:

- [OpenAI](https://openai.com): For standards that offer consensus in chaos.
- The [Rust](https://www.rust-lang.org) community for their excellent tools and libraries
- Google's [Gemma](https://deepmind.google/models/gemma) team for [gemma-3-1b-it](https://huggingface.co/google/gemma-3-1b-it)

### Open source projects that have inspired and enabled this work
- **[GenAIScript](https://github.com/microsoft/genaiscript)**: Automatable GenAI Scripting
- **[axum](https://github.com/tokio-rs/axum)**: Web framework for building APIs
- **[tokio](https://github.com/tokio-rs/tokio)**: Asynchronous runtime for efficient concurrency
- **[serde](https://github.com/serde-rs/serde)**: Serialization/deserialization framework
- **[rmcp](https://github.com/model-context-protocol/rmcp)**: Model Context Protocol SDK for agent communication
- **[sled](https://github.com/spacejam/sled)**: Embedded database for persistent storage
- **[tower-http](https://github.com/tower-rs/tower-http)**: HTTP middleware components
- **[candle-core](https://github.com/huggingface/candle)**: ML framework for efficient tensor operations
- **[candle-transformers](https://github.com/huggingface/candle/tree/main/candle-transformers)**: Implementation of
  transformer models in Candle
- **[hf-hub](https://github.com/huggingface/hf-hub)**: Client for downloading models from Hugging Face
- **[tokenizers](https://github.com/huggingface/tokenizers)**: Fast text tokenization for ML models
- **[safetensors](https://github.com/huggingface/safetensors)**: Secure format for storing tensors

## Architecture Diagram

```mermaid
%% High‑fidelity architecture diagram – client‑ready
flowchart LR
    %% ─────────────── Agent‑side ───────────────
    subgraph AGENT_SERVER["Agent Server"]
        direction TB
        AS["Agent Server"]:::core -->|exposes| MCP[["Model Context Protocol API"]]:::api
        AS -->|serves| UI["MCP Inspector UI"]:::ui

        subgraph AGENTS["Agents"]
            direction TB
            A_SEARCH["Search Agent"] -->|uses| SEARX
            A_NEWS["News Agent"]   -->|uses| SEARX
            A_SCRAPE["Web Scrape Agent"] -->|uses| BROWSER
            A_IMG["Image Generator Agent"]-->|uses| EXTERNAL_API
            A_RESEARCH["Deep Research Agent"] -->|leverages| SEARX
        end

        %% Individual fan‑out lines (no “&”)
        MCP -->|routes| A_SEARCH
        MCP -->|routes| A_NEWS
        MCP -->|routes| A_SCRAPE
        MCP -->|routes| A_IMG
        MCP -->|routes| A_RESEARCH
    end

    %% ─────────────── Local inference ───────────────
    subgraph INFERENCE["Inference Engine"]
        direction TB
        LIE["Inference Engine"]:::core -->|loads| MODELS["Gemma Models"]:::model
        LIE -->|exposes| OPENAI_API["OpenAI‑compatible API"]:::api
        MODELS -->|runs on| ACCEL

        subgraph ACCEL["Hardware Acceleration"]
            direction LR
            METAL[Metal]
            CUDA[CUDA]
            CPU[CPU]
        end
    end

    %% ─────────────── External bits ───────────────
    subgraph EXTERNAL["External Components"]
        direction TB
        SEARX["SearXNG Search"]:::ext
        BROWSER["Chromium Browser"]:::ext
        EXTERNAL_API["Public OpenAI API"]:::ext
    end

    %% ─────────────── Clients ───────────────
    subgraph CLIENTS["Client Applications"]
        CLIENT["MCP‑aware Apps"]:::client
    end

    %% ─────────────── Interactions ───────────────
    CLIENT -- "HTTPS / WebSocket" --> MCP
    AS --> |"may call"| OPENAI_API
    AS --> |"optional"| EXTERNAL_API

    %% ─────────────── Styling ───────────────
    classDef core   fill:#A9CEF4,stroke:#36494E,stroke-width:2px,color:#000;
    classDef api    fill:#7EA0B7,stroke:#36494E,stroke-width:2px,color:#000;
    classDef ui     fill:#A9CEF4,stroke:#597081,stroke-dasharray:4 3,color:#000;
    classDef model  fill:#A9CEF4,stroke:#36494E,stroke-width:2px,color:#000;
    classDef ext    fill:#B5D999,stroke:#36494E,stroke-width:2px,color:#000;
    classDef client fill:#FFE69A,stroke:#36494E,stroke-width:2px,color:#000;

```

## Setup

1. Clone the repository
2. Copy the example environment file:
   ```bash
   cp .env.example .env
   ```
3. Install JavaScript dependencies:
   ```bash
   bun i
   ```
4. Start the SearXNG search engine:
   ```bash
   docker compose up -d searxng
   ```

## Running the Project

### Inference Engine

To run the local inference engine:

```bash
cd crates/inference-engine
cargo run --release -- --server
```

### Agent Server

To run the agent server:

```bash
cargo run -p agent-server
```

### Development Mode

For development with automatic reloading:

```bash
bun dev
```

## Building

To build all crates in the workspace:

```bash
cargo build
```

To build a specific crate:

```bash
cargo build -p agent-server
# or
cargo build -p inference-engine
```