# Filesystem MCP Assistant

This project upgrades the original milestone 1 filesystem helper approach to an MCP-based architecture.

## What changed

- The filesystem business logic now lives in a shared domain layer.
- `mcp_filesystem_server.py` exposes the functionality through MCP-style JSON-RPC methods.
- `llm_file_assistant.py` talks to the filesystem through an MCP client wrapper instead of importing direct helpers.
- Batch processing and watch sessions are exposed as MCP tools.
- Resource discovery is exposed through MCP resources and resource templates.

## Setup

```powershell
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
```

Set your model credentials in `.env` or environment variables:

```env
OPENROUTER_API_KEY=your_key_here
OPENROUTER_MODEL=openai/gpt-4o-mini
OPENROUTER_HTTP_REFERER=http://localhost
OPENROUTER_APP_NAME=Filesystem MCP Assistant
```

## Run

Generate sample resumes:

```powershell
python generate_mock_data.py
```

Start the agent:

```powershell
python llm_file_assistant.py
```

Run the test suite:

```powershell
pytest
```

## MCP workflow

```mermaid
stateDiagram-v2
    [*] --> Agent
    Agent --> MCPClient: list tools/resources
    MCPClient --> MCPServer: initialize + JSON-RPC request
    MCPServer --> MCPClient: tool/resource result
    MCPClient --> Agent: structured result
    Agent --> Model: tool message + context
    Model --> Agent: final answer or another tool call
    Agent --> [*]: task complete
```

## Test scenarios

- Resource discovery returns capabilities and resource templates.
- `watch_directory` reports new resume files after a second poll.
- `batch_process` can read or search multiple resumes in one call.
- The agent can resolve a tool call through MCP and produce a final response.
- The end-to-end flow works with a real MCP subprocess and a stubbed model response.