code-mode
๐ Plug-and-play library to enable agents to call MCP and UTCP tools via code execution.
Documentation
> Transform your AI agents from clunky tool callers into efficient code executors โ in just 3 lines.
Why This Changes Everything
LLMs excel at writing code but struggle with tool calls. Instead of exposing hundreds of tools directly, give them ONE tool that executes TypeScript code with access to your entire toolkit.
Apple, Cloudflare, and Anthropic say that Code-Mode is a more efficient way to approach tool calling compared to the traditional dump function information and then extract a JSON for function calling.
Benchmarks
Independent Python benchmark study validates the performance claims with $9,536/year cost savings at 1,000 scenarios/day:
| Scenario Complexity | Traditional | Code Mode | Improvement |
|---|---|---|---|
| Simple (2-3 tools) | 3 iterations | 1 execution | 67% faster |
| Medium (4-7 tools) | 8 iterations | 1 execution | 75% faster |
| Complex (8+ tools) | 16 iterations | 1 execution | 88% faster |
Why Code Mode Dominates:
Batching Advantage - Single code block replaces multiple API calls
Cognitive Efficiency - LLMs excel at code generation vs. tool orchestration
Computational Efficiency - No context re-processing between operations
Getting Started
[
](https://www.youtube.com/watch?v=zsMjkPzmqhA)
Get Started in 3 Lines
import { CodeModeUtcpClient } from '@utcp/code-mode';
const client = await CodeModeUtcpClient.create(); // 1. Initialize
await client.registerManual({ name: 'github', /* MCP config */ }); // 2. Add tools
const { result } = await client.callToolChain(`/* TypeScript */`); // 3. Execute codeThat's it. Your AI agent can now execute complex workflows in a single request instead of dozens.
What You Get
Progressive Tool Discovery
// Agent discovers tools dynamically, loads only what it needs
const tools = await client.searchTools('github pull request');
// Instead of 500 tool definitions โ 3 relevant toolsNatural Code Execution
const { result, logs } = await client.callToolChain(`
// Chain multiple operations in one request
const pr = await github.get_pull_request({ owner: 'microsoft', repo: 'vscode', pull_number: 1234 });
const comments = await github.get_pull_request_comments({ owner: 'microsoft', repo: 'vscode', pull_number: 1234 });
const reviews = await github.get_pull_request_reviews({ owner: 'microsoft', repo: 'vscode', pull_number: 1234 });
// Process data efficiently in-sandbox
return {
title: pr.title,
commentCount: comments.length,
approvals: reviews.filter(r => r.state === 'APPROVED').length
};
`);
// Single API call replaces 15+ traditional tool callsAuto-Generated TypeScript Interfaces
namespace github {
interface get_pull_requestInput {
/** Repository owner */
owner: string;
/** Repository name */
repo: string;
/** Pull request number */
pull_number: number;
}
}Enterprise-Ready
- Secure VM Sandboxing โ Node.js isolates prevent unauthorized access
- Timeout Protection โ Configurable execution limits prevent runaway code
- Complete Observability โ Full console output capture and error handling
- Zero External Dependencies โ Tools only accessible through registered UTCP/MCP servers
- Runtime Introspection โ Dynamic interface discovery for adaptive workflows
If you're working at an enterprise, and need support, book a consultation here.
Universal Protocol Support
Works with any tool ecosystem:
| Protocol | Description | Usage |
|---|---|---|
| MCP | Model Context Protocol servers | `call_template_type: 'mcp'` |
| HTTP | REST APIs with auto-discovery | `call_template_type: 'http'` |
| File | Local JSON/YAML configurations | `call_template_type: 'file'` |
| CLI | Command-line tool execution | `call_template_type: 'cli'` |
Installation
npm install @utcp/code-modeRecommended for shell agents: the `utcp` CLI
If your agent can run shell commands (Claude Code, Cursor, Codex, Claude Cowork, etc.), the `utcp` CLI is the preferred way to use Code Mode โ no MCP server, no client config, no env vars. The agent self-configures by writing a `.utcp_config.json` and drives everything from the shell.
Just point the agent at the built-in guide:
npx -y @utcp/code-mode-cli promptHand it a description of the API to use โ a UTCP call template, an OpenAPI spec, or a plain-English description โ and it writes the config, discovers tools, runs tool-chains, and even completes interactive OAuth logins (e.g. Notion), all from the shell:
npx -y @utcp/code-mode-cli search "" # discover tools + TS interfaces
npx -y @utcp/code-mode-cli run b.title);
EOF
npx -y @utcp/code-mode-cli login # interactive OAuth, writes token to .env> CLI vs MCP: prefer the CLI whenever the agent has shell access (most coding agents) โ it's simpler and self-configuring. Use the MCP server (below) only for MCP-only clients like Claude Desktop. Both wrap the same `@utcp/code-mode` engine.
See `code-mode-cli/` for full docs.
Ready-to-Use MCP Server
On an MCP-only client (e.g. Claude Desktop)? Use our plug-and-play MCP server. (If your agent has a shell, prefer the `utcp` CLI above.)
{
"mcpServers": {
"code-mode": {
"command": "npx",
"args": ["@utcp/code-mode-mcp"],
"env": {
"UTCP_CONFIG_FILE": "/path/to/your/.utcp_config.json"
}
}
}
}That's it! No installation, no Node.js knowledge required. The Code Mode MCP Server automatically:
- Downloads and runs the latest version via `npx`
- Loads your tool configurations from JSON
- Provides code execution capabilities to Claude Desktop
- Gives you `call_tool_chain` as an MCP tool for TypeScript execution
Perfect for non-developers who want Code Mode power in Claude Desktop!
Direct TypeScript Usage
1. MCP Server Integration
Connect to any Model Context Protocol server:
> Each `call_template_type` is a separate plugin package. Install and import the
> package once at startup so it can register itself with the client; the plugin
> registers as a side effect of being imported. For `mcp`, that's `@utcp/mcp`.
> The same pattern applies to other transports: `@utcp/http`, `@utcp/text`, etc.
>
> ```bash
> npm install @utcp/mcp
> ```
import '@utcp/mcp'; // registers the 'mcp' call template
import { CodeModeUtcpClient } from '@utcp/code-mode';
const client = await CodeModeUtcpClient.create();
// Connect to GitHub MCP server
await client.registerManual({
name: 'github',
call_template_type: 'mcp',
config: {
mcpServers: {
github: {
transport: 'stdio', // required by @utcp/mcp
command: 'docker',
args: ['run', '-i', '--rm', '-e', 'GITHUB_PERSONAL_ACCESS_TOKEN', 'mcp/github'],
env: { GITHUB_PERSONAL_ACCESS_TOKEN: process.env.GITHUB_TOKEN }
}
}
}
});2. Execute Multi-Step Workflows
Replace 15+ tool calls with a single code execution:
const { result, logs } = await client.callToolChain(`
// Traditional: 4 separate API round trips โ Code Mode: 1 execution
const pr = await github.get_pull_request({ owner: 'microsoft', repo: 'vscode', pull_number: 1234 });
const comments = await github.get_pull_request_comments({ owner: 'microsoft', repo: 'vscode', pull_number: 1234 });
const reviews = await github.get_pull_request_reviews({ owner: 'microsoft', repo: 'vscode', pull_number: 1234 });
const files = await github.get_pull_request_files({ owner: 'microsoft', repo: 'vscode', pull_number: 1234 });
// Process data in-sandbox (no token overhead)
const summary = {
title: pr.title,
state: pr.state,
author: pr.user.login,
stats: {
comments: comments.length,
reviews: reviews.length,
filesChanged: files.length,
approvals: reviews.filter(r => r.state === 'APPROVED').length
},
topDiscussion: comments.slice(0, 3).map(c => ({
author: c.user.login,
preview: c.body.substring(0, 100) + '...'
}))
};
console.log(\`PR "\${pr.title}" analysis complete\`);
return summary;
`);
console.log('Analysis Result:', result);
// console output: 'PR "Fix memory leak in hooks" analysis complete'Advanced Features
Multi-Protocol Tool Chains
Mix and match different tool ecosystems in a single execution:
// Register multiple tool sources
await client.registerManual({ name: 'github', call_template_type: 'mcp', /* config */ });
await client.registerManual({ name: 'slack', call_template_type: 'http', /* config */ });
await client.registerManual({ name: 'db', call_template_type: 'file', file_path: './db-tools.json' }); // This loads a UTCP manual from a json file
const result = await client.callToolChain(`
// Fetch PR data from GitHub (MCP)
const pr = await github.get_pull_request({ owner: 'company', repo: 'api', pull_number: 42 });
// Query deployment status from database (File)
const deployment = await db.get_deployment_status({ pr_id: pr.id });
// Send notification to Slack (HTTP)
await slack.post_message({
channel: '#releases',
text: \`PR #42 "\${pr.title}" deployed to \${deployment.environment}\`
});
return { pr: pr.title, environment: deployment.environment };
`);Runtime Interface Introspection
Tools can dynamically discover and adapt to available interfaces:
const result = await client.callToolChain(`
// Discover available tools at runtime
console.log('Available interfaces:', __interfaces);
// Get specific tool interface for validation
const prInterface = __getToolInterface('github.get_pull_request');
console.log('PR tool expects:', prInterface);
// Use interface info for dynamic workflows
const hasSlackTools = __interfaces.includes('namespace slack');
if (hasSlackTools) {
await slack.post_message({ channel: '#dev', text: 'Analysis complete' });
}
return { toolsAvailable: hasSlackTools };
`);Context-Efficient Data Processing
Process large datasets without bloating the model's context:
const result = await client.callToolChain(`
// Fetch large dataset
const allIssues = await github.list_repository_issues({ owner: 'facebook', repo: 'react' });
console.log('Fetched', allIssues.length, 'total issues');
// Process efficiently in-sandbox
const criticalBugs = allIssues
.filter(issue => issue.labels.some(l => l.name === 'bug'))
.filter(issue => issue.labels.some(l => l.name === 'high priority'))
.map(issue => ({
number: issue.number,
title: issue.title,
author: issue.user.login,
daysOld: Math.floor((Date.now() - new Date(issue.created_at)) / (1000 * 60 * 60 * 24))
}))
.sort((a, b) => b.daysOld - a.daysOld);
// Only return processed summary (not 10,000 raw issues)
return {
totalIssues: allIssues.length,
criticalBugs: criticalBugs.slice(0, 10), // Top 10 oldest critical bugs
summary: \`Found \${criticalBugs.length} critical bugs, oldest is \${criticalBugs[0]?.daysOld} days old\`
};
`);Error Handling & Observability
Built-in error handling with complete execution transparency:
const { result, logs } = await client.callToolChain(`
try {
console.log('Starting multi-step workflow...');
const data = await external_api.fetch_data({ id: 'user-123' });
console.log('Data fetched successfully');
const processed = await data_processor.transform(data);
console.warn('Processing completed with', processed.warnings.length, 'warnings');
return processed;
} catch (error) {
console.error('Workflow failed:', error.message);
throw error; // Propagates to outer error handling
}
`, 30000); // 30-second timeout
// Complete observability
console.log('Result:', result);
console.log('Execution logs:', logs);
// ['Starting multi-step workflow...', 'Data fetched successfully', '[WARN] Processing completed with 2 warnings']Custom Timeouts
Configure execution limits for different workload types:
// Quick operations (5 seconds)
const quickResult = await client.callToolChain(`return await ping.check();`, 5000);
// Heavy data processing (2 minutes)
const heavyResult = await client.callToolChain(`
const bigData = await database.export_full_dataset();
return await analytics.process_dataset(bigData);
`, 120000);AI Agent Integration
Plug-and-play with any AI framework. The built-in prompt template handles all the complexity:
import { CodeModeUtcpClient } from '@utcp/code-mode';
const systemPrompt = `
You are an AI assistant with access to tools via UTCP CodeMode.
${CodeModeUtcpClient.AGENT_PROMPT_TEMPLATE}
Additional instructions...
`;
// Works with any AI library
const response = await openai.chat.completions.create({
model: 'gpt-4',
messages: [
{ role: 'system', content: systemPrompt },
{ role: 'user', content: 'Analyze the latest PR in microsoft/vscode' }
]
});The template provides comprehensive guidance on:
- Tool discovery workflow (`searchTools` โ `__interfaces` โ `callToolChain`)
- Hierarchical access patterns (`manual.tool()` syntax)
- Interface introspection (`__getToolInterface()`)
- Error handling and best practices
API Reference
Core Methods
`callToolChain(code: string, timeout?: number)`
Execute TypeScript code with full tool access and observability.
- Returns: `{result: any, logs: string[]}` with execution result and captured console output
- Default timeout: 30 seconds
`getAllToolsTypeScriptInterfaces()`
Generate complete TypeScript interfaces for IDE integration.
- Returns: String containing all interface definitions with namespaces
`searchTools(query: string)` *(from UtcpClient)*
Discover tools using natural language queries.
- Returns: Array of relevant tools with descriptions and interfaces
Static Methods
`CodeModeUtcpClient.create(root_dir?, config?)`
Create a new client instance with optional configuration.
`CodeModeUtcpClient.AGENT_PROMPT_TEMPLATE`
Production-ready prompt template for AI agents.
Security & Performance
Secure by Design
- Node.js VM sandboxing โ Isolated execution context
- No filesystem access โ Tools only through registered servers
- Timeout protection โ Configurable execution limits
- Zero network access โ No external dependencies or API keys exposed
Performance Optimized
- Minimal memory footprint โ VM contexts are lightweight
- Efficient tool caching โ TypeScript interfaces cached automatically
- Streaming console output โ Real-time log capture without buffering
- Identifier sanitization โ Handles invalid TypeScript identifiers gracefully
Development Experience
IDE Integration
Generate TypeScript definitions for full IntelliSense support:
# Generate tool interfaces
const interfaces = await client.getAllToolsTypeScriptInterfaces();
await fs.writeFile('generated-tools.d.ts', interfaces);
# Add to tsconfig.json
{
"compilerOptions": {
"typeRoots": ["./generated-tools.d.ts"]
}
}Debug & Monitor
Built-in observability for production deployments:
const { result, logs } = await client.callToolChain(userCode);
// Ship logs to your monitoring system
logs.forEach(log => {
if (log.startsWith('[ERROR]')) monitoring.error(log);
if (log.startsWith('[WARN]')) monitoring.warn(log);
});Benchmark Methodology
The comprehensive Python study tested 16 realistic scenarios across:
- Financial workflows (invoicing, expense tracking)
- DevOps operations (deployments, monitoring)
- Data processing (analysis, reporting)
- Business automation (CRM, notifications)
Models tested: Claude Haiku, Gemini Flash
Pricing basis: $0.25/1M input, $1.25/1M output tokens
Scale: 1,000 scenarios/day = $9,536/year savings with Code Mode
Learn More
- **Cloudflare Research** โ Original code mode whitepaper
- **Anthropic Study** โ MCP code execution benefits
- **Python Benchmark Study** โ Comprehensive performance analysis
- **UTCP Specification** โ Official TypeScript implementation
- **Report Issues** โ Bug reports and feature requests
License
MPL-2.0 โ Open source with commercial-friendly terms.
Frequently asked questions
What is code-mode?
code-mode is ๐ Plug-and-play library to enable agents to call MCP and UTCP tools via code execution.
How do I install code-mode?
Open the GitHub repository and follow its README. Most MCP servers are added to your client's MCP config, then called by your agent.
Is code-mode open source?
Yes โ it is hosted on GitHub at https://github.com/universal-tool-calling-protocol/code-mode and has 1,548 stars.
Related MCP tools
A Model Context Protocol server for converting almost anything to Markdown TypeScript-based implementation. Trusted by 2200+ developers.
Your memories are in ChatGPT... But nowhere else. Universal Memory MCP makes your memories available to every single LLM. No logins or paywall. One command to set it up.
MCP Server for kubernetes management commands
MCP Aggregator, Orchestrator, Middleware, Gateway in one docker TypeScript-based implementation. Trusted by 1400+ developers.
Composio equips your AI agents & LLMs with 100+ high-quality integrations via function calling for the Model Context Protocol. Enhance AI assistants with powerf
A MCP for Claude Desktop / Claude Code / Windsurf / Cursor to build n8n workflows for you
Run your own MCP server? See who uses it and what to fix.
Measure it with TrackMCP