Production-Ready Code Execution Platform for AI Agents
A high-performance, MCP-enabled code execution service that lets AI agents run JavaScript/TypeScript with full access to tools, APIs, and knowledge systems. Built for speed, security, and seamless integration with Claude Code and other AI platforms.
- MCP Server Mode: Direct integration with Claude Code via Model Context Protocol
- HTTP Backend Mode: RESTful API for any AI platform or custom integration
- Both architectures share the same powerful runtime engine
- Bun Runtime ⭐ Recommended - Full async/await, modern APIs, fastest execution
- QuickJS - Lightweight, 5-10ms startup, perfect for quick operations
- Deno - Secure by default with explicit permissions
- isolated-vm - V8 isolates for maximum control
- E2B - Cloud sandboxes for untrusted code
Execute code with instant access to your entire MCP ecosystem:
- AutoMem: AI memory with graph relationships and vector search
- Context7: Official library documentation on demand
- Sequential Thinking: Multi-step reasoning for complex analysis
- HelpScout: Customer support data access
- WordPress API: Content management operations
- ...and any other MCP server you have configured
- Sandboxed execution with memory/CPU limits
- Capability-based permissions
- Network access controls
- Audit logging for all operations
- OAuth 2.1 + JWT authentication
- Node.js 20+
- npm or yarn
- (Optional) Bun runtime for best performance
# Clone or navigate to the project
cd codemode-unified
# Install dependencies
npm install
# Build the project
npm run buildOption 1: MCP Server (for Claude Code)
npm run start:mcp
# MCP server starts on stdio - ready for Claude Code integrationOption 2: HTTP Backend
npm start
# HTTP server starts on http://localhost:3001Option 3: Development Mode
npm run dev # Auto-reload on changes
npm run dev:bun # Use Bun runtime specificallyAdd to your .mcp.json:
{
"mcpServers": {
"codemode-unified": {
"command": "node",
"args": ["/path/to/codemode-unified/dist/mcp-server.js"],
"env": {
"MCP_CONFIG_PATH": "/path/to/your/.mcp.json"
}
}
}
}The MCP server automatically discovers and connects to all your other MCP servers (AutoMem, Context7, etc.) from the same config file.
Create .env:
PORT=3001
HOST=localhost
# MCP Integration
MCP_CONFIG_PATH=/path/to/your/.mcp.json
# Optional: Runtime selection
DEFAULT_RUNTIME=bun// Claude can execute this directly through the MCP tool
const pokemonResponse = await fetch('https://pokeapi.co/api/v2/pokemon/pikachu');
const pokemon = await pokemonResponse.json();
// Store in AutoMem for later recall
const memory = await mcp.automem.store_memory({
content: `Pokemon: ${pokemon.name} has ${pokemon.abilities.length} abilities`,
tags: ['pokemon', 'api-test'],
importance: 0.8
});
// Return structured results
return {
pokemon: pokemon.name,
memory_id: memory.memory_id,
status: 'success'
};curl -X POST http://localhost:3001/execute \
-H "Content-Type: application/json" \
-d '{
"code": "const result = await fetch(\"https://api.github.com/users/danieliser\").then(r => r.json()); return { username: result.login, repos: result.public_repos };",
"runtime": "bun",
"timeout": 10000
}'Response:
{
"success": true,
"result": {
"username": "danieliser",
"repos": 42
},
"stdout": "",
"stderr": "",
"execution_time_ms": 261
}// Fetch data from multiple APIs in parallel
const [github, twitter, linkedin] = await Promise.all([
fetch('https://api.github.com/users/username').then(r => r.json()),
fetch('https://api.twitter.com/user/username').then(r => r.json()),
fetch('https://api.linkedin.com/profile/username').then(r => r.json())
]);
// Store aggregated profile in AutoMem
await mcp.automem.store_memory({
content: `Profile: ${github.name} - ${github.bio}`,
tags: ['profile', 'social-media'],
metadata: {
github_repos: github.public_repos,
twitter_followers: twitter.followers_count,
linkedin_connections: linkedin.connections
}
});
return { profile: 'stored', sources: 3 };// Use Context7 to get official React docs
const reactDocs = await mcp.context7.get_library_docs({
context7CompatibleLibraryID: '/facebook/react',
topic: 'hooks'
});
// Analyze with Sequential Thinking
const analysis = await mcp.sequential_thinking.sequentialthinking({
thought: "Compare React hooks patterns with Vue Composition API",
thoughtNumber: 1,
totalThoughts: 3,
nextThoughtNeeded: true
});
return { documentation: reactDocs, analysis };// Fetch related entities
const company = await fetch(`https://api.example.com/companies/${id}`).then(r => r.json());
// Store as memory nodes
const companyMemory = await mcp.automem.store_memory({
content: `Company: ${company.name} in ${company.industry}`,
tags: ['company', company.industry]
});
// Create relationships
for (const product of company.products) {
const productMemory = await mcp.automem.store_memory({
content: `Product: ${product.name}`,
tags: ['product']
});
await mcp.automem.associate_memories({
memory1_id: companyMemory.memory_id,
memory2_id: productMemory.memory_id,
type: 'RELATES_TO',
strength: 0.9
});
}
return { graph_nodes: company.products.length + 1 };┌─────────────────────────────────────────────────────────┐
│ CodeMode Unified │
├─────────────────────┬───────────────────────────────────┤
│ │ │
│ MCP Server │ HTTP Backend │
│ (stdio) │ (port 3001) │
│ │ │
│ ├─ MCP Client │ ├─ REST API │
│ ├─ Tool Bridge │ ├─ WebSocket │
│ └─ Executor │ └─ MCP Aggregator │
│ │ │ │ │
└────────┼────────────┴───────────┼───────────────────────┘
│ │
└────────────┬───────────┘
▼
┌──────────────────────────┐
│ Runtime Executor │
├──────────────────────────┤
│ ├─ Bun Runtime │
│ ├─ QuickJS Runtime │
│ ├─ Deno Runtime │
│ ├─ Isolated-VM │
│ └─ E2B Runtime │
└──────────────────────────┘
│
┌────────────┴─────────────┐
│ │
▼ ▼
┌─────────┐ ┌─────────────┐
│ MCP │ │ Standard │
│ Tools │ │ APIs │
└─────────┘ └─────────────┘
AutoMem fetch()
Context7 console
Sequential Date, JSON, etc
- executor.ts - Core orchestration and runtime selection
- mcp-server.ts - MCP protocol implementation with tool bridge
- server.ts - HTTP/WebSocket server with REST API
- runtime/bun-runtime.ts - Bun execution engine (recommended)
- mcp-aggregator.ts - Multi-server MCP client
# Run all tests
npm test
# Test specific runtimes
npm run test:runtime:bun
npm run test:runtime:quickjs
npm run test:runtime:deno
# Coverage report
npm run test:coverage# Run all tests
npm test
# Test specific runtimes
npm run test:runtime:bun
npm run test:runtime:quickjs
npm run test:runtime:deno
# Coverage report
npm run test:coverageComprehensive benchmark suite available: examples/benchmarks/
# Quick comparison (Deno vs Bun)
PATH="/Users/$USER/.deno/bin:$PATH" node examples/benchmarks/compare-runtimes.js
# Full benchmark suite (3 runtimes, 20+ tests)
PATH="/Users/$USER/.deno/bin:$PATH" node examples/benchmarks/runtime-comparison.js
# Deno validation suite
PATH="/Users/$USER/.deno/bin:$PATH" node examples/benchmarks/test-deno-runtime.js
# Real-world workflow test
PATH="/Users/$USER/.deno/bin:$PATH" node examples/benchmarks/test-deno-pokemon.jsBenchmark Results (Actual Testing):
| Test Type | Deno | Bun | Winner |
|---|---|---|---|
| Simple Expression | 21ms | 74ms | 🦕 Deno |
| Array Operations | 19ms | 10ms | 🍞 Bun |
| Single Fetch | 582ms | 183ms | 🍞 Bun |
| Parallel Fetches | 353ms | 152ms | 🍞 Bun |
| Average | 199ms | 86ms | 🍞 Bun (2.3x faster) |
See examples/benchmarks/README.md for complete analysis.
✅ 6/6 Workflows Passed (100% success rate)
Tested Capabilities:
- ✅ Multi-API data aggregation
- ✅ Async/await with Promise.all
- ✅ MCP tool integration (store, recall, associate)
- ✅ Structured object returns
- ✅ Knowledge graph creation
- ✅ Error handling and recovery
- ✅ Complex nested data structures
- ✅ Real-world API interactions
Performance:
- Execution Time: 150-300ms per workflow (Bun), 350-1400ms (Deno)
- Memory Usage: <50MB per sandbox
- Throughput: 1000+ requests/second (Bun), 500+ req/sec (Deno)
Each runtime has different strengths:
| Runtime | Async | NPM Packages | Startup | Best For |
|---|---|---|---|---|
| Bun ⭐ | ✅ Full | ✅ Yes | 50-100ms | Production workflows |
| QuickJS | ❌ No | 5-10ms | Quick computations | |
| Deno | ✅ Full | ✅ Yes | 100-200ms | Secure execution |
| isolated-vm | ✅ Full | 50-100ms | Custom isolation | |
| E2B | ✅ Full | ✅ Yes | 1-5s | Untrusted code |
All MCP tools are automatically namespaced and available:
// AutoMem (7 tools)
await mcp.automem.store_memory({...})
await mcp.automem.recall_memory({...})
await mcp.automem.associate_memories({...})
await mcp.automem.update_memory({...})
await mcp.automem.delete_memory({...})
await mcp.automem.search_memory_by_tag({...})
await mcp.automem.check_database_health()
// Context7 (2 tools)
await mcp.context7.resolve_library_id({...})
await mcp.context7.get_library_docs({...})
// Sequential Thinking (1 tool)
await mcp.sequential_thinking.sequentialthinking({...})The system intelligently parses both structured and formatted responses:
// MCP tools might return formatted text or JSON
const result = await mcp.automem.recall_memory({query: 'test'});
// Automatically parsed to structured data:
// {
// _text: "original formatted text",
// memory_id: "abc123",
// content: "stored content",
// tags: "tag1, tag2"
// }- Return Statement Wrapping - Multi-line return statements now work correctly
- MCP Response Parsing - Handles both JSON and formatted text responses
- Config Loading - Properly maps mcpServers→servers format
- Process Cleanup - No zombie processes after restarts
- Context7 Parsing - Occasional inconsistencies with complex doc structures
- Tag-based Recall - Less reliable than query-based memory recall
- QuickJS Async - Two-pass system for async emulation (use Bun for real async)
Bun Runtime: 50-100ms startup + execution time
QuickJS Runtime: 5-10ms startup + execution time
Deno Runtime: 100-200ms startup + execution time
Single-threaded: 1000+ req/sec
Worker pool (4): 3500+ req/sec
With MCP tools: 500+ req/sec (network-bound)
Base runtime: 10-20MB
Per sandbox: 5-50MB (depending on code)
MCP connections: 5-10MB per server
- ✅ Memory limits enforced (default 128MB)
- ✅ CPU quota restrictions (default 50%)
- ✅ Execution timeout (default 30s)
- ✅ Network access control
- ✅ Filesystem isolation
- JWT with HS256/RS256
- OAuth 2.1 with PKCE
- API key authentication
- Capability-based permissions
All code execution is logged with:
- Timestamp and user identity
- Code executed (sanitized)
- Runtime and resources used
- MCP tools accessed
- Success/failure status
- ARCHITECTURE.md - System design and components
- TECHNICAL_SPECIFICATION.md - Detailed specs
- RUNTIME_COMPARISON.md - Runtime trade-offs
- MCP_INTEGRATION_COMPLETE.md - MCP setup guide
- TEST_RESULTS.md - Full test reports
Contributions welcome! Please:
- Fork the repository
- Create a feature branch
- Add tests for new functionality
- Ensure
npm run typecheckpasses - Submit a pull request
MIT License - see LICENSE file for details.
Built with:
- QuickJS - Lightweight JavaScript engine
- Bun - Fast all-in-one JavaScript runtime
- Model Context Protocol - AI-native tool protocol
- Fastify - High-performance web framework
- Issues: GitHub Issues
- Discussions: GitHub Discussions
- Twitter: @danieliser
Built with ❤️ for the AI agent ecosystem