domdomegg/programmatic-mcp-prototype

An MCP-based agent with support for:
- progressive tool discovery
- programmatic tool composition
- state persistence
- skill building
Architecture
- **Core Agent Loop**: Simple while loop that can be swapped with other implementations
- **MCP Proxy Server**: Aggregates multiple MCP servers into one unified interface
- **Code Generator**: Creates TypeScript bindings from MCP tool schemas
- **Container Runner**: Executes TypeScript code in isolated Docker containers
Key Features
1. Progressive Tool Discovery
The model can search for and discover tools dynamically instead of loading all tools upfront. Rather than exposing hundreds of tools at once, the agent provides search_tools and execute_tool meta-tools that allow the model to find relevant tools as needed, reducing context usage and improving response quality.
2. Programmatic Tool Composition
The agent can write TypeScript code that composes MCP tools together:
3. State Persistence
Store intermediate results and data in the workspace directory:
4. Skill Building
Create reusable meta-tools that combine multiple operations:
Setup
1. Install dependencies:
2. Configure your MCP servers in config/servers.ts
3. Set your Anthropic API key:
4. Build Docker image for code execution:
Usage
Run the agent:
How It Works
1. **Startup**: Connects to configured MCP servers (bash, computer, container)
2. **Code Generation**: Creates TypeScript functions for each tool with proper types
3. **Agent Loop**: Simple while loop that calls Claude with MCP tools
4. **Tool Execution**: Routes tool calls to appropriate backend MCP servers
5. **Code Execution**: Runs TypeScript in isolated Docker containers
Benefits of Programmatic Tool Use
1. **Composition**: Chain multiple tools without waiting between calls
2. **State**: Store variables and reuse results
3. **Loops/Conditionals**: Handle complex logic in code
4. **Error Handling**: Try/catch and retry logic