Build self-improving agents with programmatic tool discovery

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:

// Example: The model can write code like this
import * as bash from './generated/servers/bash';
import * as computer from './generated/servers/computer';

const files = await bash.ls({ path: './documents' });
for (const file of files) {
  const content = await bash.readFile({ path: file });
  console.log(`File ${file}: ${content.length} bytes`);
}

3. State Persistence

Store intermediate results and data in the workspace directory:

import * as fs from 'fs/promises';

// Save CSV for later use
const csvData = await processData();
await fs.writeFile('./generated/workspace/data.csv', csvData);

// Load it in a future execution
const data = await fs.readFile('./generated/workspace/data.csv', 'utf-8');

4. Skill Building

Create reusable meta-tools that combine multiple operations:

// Build a skill in ./generated/skills/
export async function saveSheetAsCsv(sheetId: string) {
  import * as sheets from '../servers/sheets';
  import * as bash from '../servers/bash';
  
  const data = await sheets.getCells({ sheetId });
  const csv = data.map(row => row.join(',')).join('\n');
  const path = `../workspace/sheet-${sheetId}.csv`;
  await bash.writeFile({ path, content: csv });
  return path;
}

// Use the skill later
import { saveSheetAsCsv } from './generated/skills/save-sheet-as-csv';
const csvPath = await saveSheetAsCsv('abc123');

Setup

1. Install dependencies:

npm install

2. Configure your MCP servers in config/servers.ts

3. Set your Anthropic API key:

export ANTHROPIC_API_KEY='your-key'

4. Build Docker image for code execution:

docker build -t mcp-runner:latest src/servers/container-runner

Usage

Run the agent:

npm start

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
5. Efficiency: Make many tool calls in one execution
6. Skills Library: Build reusable patterns over time