Extending GoRunRun Local AI

1. Connect an MCP server (no code)
2. Add a built-in tool
3. Add or swap a model
4. Change behavior with prompts and styles
5. Build on the API

Connect an MCP server (no code)

Any Model Context Protocol server can give the assistant new tools. Add it in Settings → Tools & MCP, or in config/mcp.json:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/YOU/Documents"],
      "policy": "confirm"
    }
  }
}

policy can be allow, confirm (ask the user each time) or deny. Tools that change things default to confirm.

Add a built-in tool

A tool is one Python class. Create orchestrator/tools/weather.py:

from orchestrator.tools.base import Tool, ToolContext, ToolResult


class Weather(Tool):
    name = "weather"
    description = "Current weather for a city."
    parameters = {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]}
    network = True          # uses the internet: labelled "internet" in Settings → Tools

    async def run(self, args: dict, ctx: ToolContext) -> ToolResult:
        report = await fetch_weather(args["city"])   # your code
        return ToolResult(f"{args['city']}: {report}", data={"city": args["city"]})

Register it in ToolRegistry (orchestrator/policy.py) and give it a default policy in config/tools.yaml. The model sees the name, description and JSON schema. Arguments are validated before run, and whatever run returns is shown to the model as untrusted data. ctx.emit can stream progress to the UI; generate_video does this.

Add or swap a model

Models are listed in config/models.yaml. To try another MLX chat model:

llms:
  my-model:
    display_name: My Model 8B
    backend: mlx-vlm
    repo: mlx-community/My-Model-8B-4bit
    est_memory_gb: 6
    capabilities: { image: true, tools: true }

Then run uv run python scripts/download_models.py --only my-model and pick it in the model menu, or make it the default under defaults.llm. GGUF models can run through llama.cpp with backend: llama-server. Speech, embedding, reranker and video models are configured in the same file.

Change behavior with prompts and styles

• config/prompts/system.md is the base system prompt.
• config/styles/*.md are the writing styles (add a file to add a style).
• config/tools.yaml sets which tools are on and their default policies.

Build on the API

The web app uses a plain HTTP API on http://127.0.0.1:8000/api. Replies stream as Server-Sent Events. For example:

make chat MSG="Summarize the plot of Hamlet in two sentences"

See scripts/chat_cli.py for a minimal client, and orchestrator/api.py for every endpoint.