Skip to content

Repository files navigation

codemode

Two MCP tools that replace hundreds. Give an AI agent your OpenAPI spec and a request handler — it discovers and calls your entire API by writing JavaScript in a sandboxed runtime.

Instead of defining individual MCP tools for every API endpoint (list-pods, create-product, get-logs, ...), CodeMode exposes just two tools:

  • search — the agent writes JS to filter your OpenAPI spec and discover endpoints
  • execute — the agent writes JS to call your API via an injected client

This is the same pattern Cloudflare uses to expose 2,500+ API endpoints through just two MCP tools, reducing context window usage by 99.9%.

Try It

Requires mise for tooling (Node.js, pnpm, Task):

git clone https://github.com/akua-dev/codemode.git
cd codemode
mise install   # installs Node 24, pnpm 10, Task
task install   # installs dependencies
task example   # runs the Petstore demo

Fetches the real Petstore OpenAPI spec from the web, then runs search + execute against a local Hono mock — no API keys needed.

Install

pnpm add @robinbraemer/codemode

# Install a sandbox runtime (at least one):
pnpm add @robinbraemer/llrt    # Native LLRT — default candidate
pnpm add isolated-vm           # V8 isolates — data-only Node.js fallback

If both are installed, the auto-selector (createExecutor) picks native LLRT first, then isolated-vm on Node for data-only execution. Request-capable execution requires LLRT. QuickJSExecutor is still exported for explicit advanced use, but it is not selected automatically because its host-callback bridge cannot enforce all byte limits before values cross into host JavaScript.

Quick Start

import { CodeMode } from '@robinbraemer/codemode';
import { Hono } from 'hono';

const app = new Hono();
app.get('/v1/clusters', (c) => c.json([{ id: '1', name: 'prod' }]));
app.post('/v1/clusters', async (c) => {
  const body = await c.req.json();
  return c.json({ id: '2', ...body }, 201);
});

const codemode = new CodeMode({
  spec: myOpenAPISpec,              // OpenAPI 3.x spec, or async getter
  request: app.request.bind(app),   // in-process, no network hop
});

// The agent searches the spec to discover endpoints...
const search = await codemode.callTool('search', {
  code: `async () => {
    const results = [];
    for (const [path, methods] of Object.entries(spec.paths)) {
      for (const [method, op] of Object.entries(methods)) {
        if (op.tags?.some(t => t.toLowerCase() === 'clusters')) {
          results.push({ method: method.toUpperCase(), path, summary: op.summary });
        }
      }
    }
    return results;
  }`
});

// ...then executes API calls
const result = await codemode.callTool('execute', {
  code: `async () => {
    const res = await api.request({ method: "GET", path: "/v1/clusters" });
    return res.body;
  }`
});

MCP Server Integration

import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js';
import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
import { CodeMode } from '@robinbraemer/codemode';
import { registerTools } from '@robinbraemer/codemode/mcp';

const codemode = new CodeMode({
  spec: () => fetchOpenAPISpec(),
  request: app.request.bind(app),
});

const server = new McpServer({ name: 'my-api', version: '1.0.0' });
registerTools(codemode, server);

const transport = new StdioServerTransport();
await server.connect(transport);

How It Works

AI Agent
  │ writes JavaScript code
  ▼
CodeMode MCP Server
  │
  ├─ search(code) → runs JS with preprocessed OpenAPI spec
  │   → ref-resolved paths view with only essential fields kept
  │   → agent discovers endpoints, schemas, parameters
  │
  └─ execute(code) → runs JS with injected request client
      → api.request() calls your handler in-process
      → no network hop, auth handled automatically

All code runs in a fresh sandbox runtime. The sandbox has zero I/O by default — no require, no process, no fetch, no filesystem. Request-capable execution uses injected host callbacks (spec for search, {namespace}.request() for execute) and is supported by LLRT.

Each tool call gets a fresh sandbox with no state carried over between calls.

API

new CodeMode(options)

Option Type Default Description
spec OpenAPISpec | () => OpenAPISpec | Promise<OpenAPISpec> required OpenAPI 3.x spec or async getter
request (input, init?) => Response required Fetch-compatible handler (app.request.bind(app) for Hono)
namespace string "api" Client name in sandbox (api.request(...)). Must be a valid JS identifier, not a reserved name.
baseUrl string "http://localhost" Base URL for relative paths
sandbox SandboxOptions see below Sandbox resource limits
executor Executor createExecutor() Custom sandbox executor
maxResponseTokens number 25000 Token limit for response truncation (0 to disable)
maxRequests number 50 Max requests per execute() call
maxResponseBytes number 10485760 Max response body size in bytes (10MB)
allowedHeaders string[] undefined Header whitelist. When unset, a blocklist strips Authorization, Cookie, Host, X-Forwarded-*, Proxy-*.
maxRefDepth number 50 Max $ref resolution depth

Successful spec preprocessing is cached. Concurrent searches share an in-flight async spec provider, while a failed preparation is retried by a later search.

Data-only input first has a 100,000-node structural inspection limit. Built-in executors additionally reject cycles, bigint, and custom toJSON hooks, then preflight the fully expanded transport shape before marshalling it into a sandbox. Shared aliases count once per transport occurrence, with limits of 500,000 expanded occurrences and 10 MiB of encoded input. Custom executors still receive the alias-preserving structured clone and remain responsible for their own transport limits.

SandboxOptions

Option Type Default Description
memoryMB number 64 Sandbox heap memory limit
timeoutMs number 30000 CPU timeout in ms (caps pure compute)
wallTimeMs number 60000 Wall-clock timeout in ms (caps total elapsed time including async I/O)

Methods

codemode.tools(): ToolDefinition[]

Returns MCP-compatible tool definitions for search and execute.

codemode.callTool(name, { code }): Promise<ToolCallResult>

Route a tool call. Returns { content: [{ type: "text", text }], isError? }.

codemode.search(code): Promise<ToolCallResult>

Run search code directly (shorthand for callTool('search', { code })).

codemode.execute(code): Promise<ToolCallResult>

Run execute code directly (shorthand for callTool('execute', { code })).

codemode.setToolNames(search, execute): this

Override default tool names. Useful when running multiple CodeMode instances.

codemode.dispose(): void

Clean up sandbox resources.

Sandbox API

Inside search

The spec global is the preprocessed OpenAPI paths view; resolvable $ref pointers are expanded inline:

// Find endpoints by tag
async () => {
  const results = [];
  for (const [path, methods] of Object.entries(spec.paths)) {
    for (const [method, op] of Object.entries(methods)) {
      if (op.tags?.some(t => t.toLowerCase() === 'clusters')) {
        results.push({ method: method.toUpperCase(), path, summary: op.summary });
      }
    }
  }
  return results;
}

// Get endpoint with requestBody schema (resolvable refs are expanded)
async () => {
  const op = spec.paths['/v1/products']?.post;
  return { summary: op?.summary, requestBody: op?.requestBody };
}

// Spec shape
async () => ({
  endpoints: Object.keys(spec.paths).length,
})

Inside execute

The {namespace}.request() function makes API calls through the host handler:

// GET with query params
async () => {
  const res = await api.request({
    method: "GET",
    path: "/v1/clusters",
    query: { limit: 10 },
  });
  return res.body;
}

// POST with body
async () => {
  const res = await api.request({
    method: "POST",
    path: "/v1/products",
    body: { name: "Redis", chart: "bitnami/redis" },
  });
  return { status: res.status, body: res.body };
}

// Chain calls
async () => {
  const list = await api.request({ method: "GET", path: "/v1/clusters" });
  const details = await Promise.all(
    list.body.map(c =>
      api.request({ method: "GET", path: `/v1/clusters/${c.id}` })
    )
  );
  return details.map(d => d.body);
}

Request options:

Field Type Description
method string HTTP method ("GET", "POST", etc.)
path string API path ("/v1/clusters")
query Record<string, string | number | boolean> Query parameters (optional)
body unknown Request body, auto-serialized as JSON (optional)
headers Record<string, string> Additional headers (optional)

Response: { status: number, headers: Record<string, string>, body: unknown }

Spec Preprocessing

CodeMode automatically preprocesses your OpenAPI spec before passing it to the search sandbox:

  • $ref resolution — resolvable $ref pointers are expanded inline; circular refs become { $circular: ref }, and refs beyond maxRefDepth become { $circular: ref, $reason: "max depth exceeded" }
  • Field extraction — only essential fields kept per operation: summary, description, tags, parameters, requestBody, responses
  • Output shape — only { paths } is passed to search; info, servers, and components are omitted because operation fields and resolved references are represented in the paths view

You can also use the preprocessing utilities directly:

import { resolveRefs, processSpec, extractTags } from '@robinbraemer/codemode';

const processed = processSpec(rawSpec);
const tags = extractTags(rawSpec);

Executors

CodeMode ships three executor backends. LlrtNativeExecutor is the default and the only request-capable backend. IsolatedVMExecutor is the Node.js data-only fallback, and QuickJSExecutor is an explicit data-only compatibility backend.

Use createExecutor() for automatic selection, or pass an executor instance explicitly:

import { CodeMode, createExecutor, IsolatedVMExecutor, LlrtNativeExecutor } from '@robinbraemer/codemode';

// Automatic — picks native LLRT first, then data-only isolated-vm on Node
const codemode = new CodeMode({
  spec,
  request: handler,
  executor: await createExecutor({ memoryMB: 128, timeoutMs: 60_000 }),
});

// Or explicit
const codemode = new CodeMode({
  spec,
  request: handler,
  executor: new IsolatedVMExecutor({
    memoryMB: 128,
    timeoutMs: 60_000,   // CPU time limit
    wallTimeMs: 120_000, // total elapsed time limit
  }),
});
Executor Package Performance Portability Production-ready
LlrtNativeExecutor @robinbraemer/llrt Lightweight native LLRT Node.js ✅ default candidate
IsolatedVMExecutor isolated-vm Native V8 speed Node.js ⚠️ data-only fallback
QuickJSExecutor quickjs-emscripten Slower (interpreted WASM) Node, Bun, CF Workers, browser ⚠️ explicit only — see caveats

QuickJSExecutor caveats

  • Not a production backend and not auto-selected. Use this only by constructing new QuickJSExecutor(...) explicitly. Production request-capable callers should use LlrtNativeExecutor.
  • Host callbacks are disabled. QuickJS cannot enforce host-call byte limits before guest values are dumped into host JavaScript, and sequential host awaits still hit upstream release-asyncify crashes. QuickJSExecutor now fails closed when function globals are provided.
  • Return-value semantics differ from isolated-vm. Final values cross via a JSON.stringify envelope. Date, Map, Set, BigInt are converted to strings/objects, not preserved as instances. isolated-vm uses structured clone and preserves them. Stick to plain JSON-safe shapes in sandboxed code that targets both backends.
  • CPU timeout is wall-clock-based. isolated-vm uses true CPU time; QuickJS uses elapsed time.

Custom Executor

Implement the Executor interface to use your own sandbox:

CodeMode passes a fresh structured clone of the processed spec to each search() call, so mutations made by a custom executor cannot change later searches.

CodeMode calls executeData() for search() and executeWithCapabilities() for execute(). Implement the latter when the custom runtime needs to expose the request capability; the legacy execute() method remains available for direct executor use.

import {
  CodeMode,
  emptyExecuteStats,
  type CapabilityManifest,
  type ExecuteResult,
  type Executor,
} from '@robinbraemer/codemode';

class MyExecutor implements Executor {
  async executeData(_code: string, _input: Record<string, unknown>): Promise<ExecuteResult> {
    // Run data-only code in the sandbox.
    return { result: undefined, stats: emptyExecuteStats() };
  }

  async executeWithCapabilities(
    _code: string,
    _input: Record<string, unknown>,
    _capabilities: CapabilityManifest,
  ): Promise<ExecuteResult> {
    // Run code with declared capabilities, such as `{namespace}.request()`.
    return { result: undefined, stats: emptyExecuteStats() };
  }

  async execute(_code: string, _globals: Record<string, unknown>): Promise<ExecuteResult> {
    // Legacy direct-executor entrypoint.
    return { result: undefined, stats: emptyExecuteStats() };
  }

  dispose() { /* clean up */ }
}

const codemode = new CodeMode({
  spec,
  request: handler,
  executor: new MyExecutor(),
});

Token Efficiency

Approach Context Tokens
Individual MCP tools (15-50+ tools) ~15,000-50,000+
Full OpenAPI spec in context ~1,000,000+
CodeMode (2 tools) ~1,000

License

MIT

About

Programmatic tool calling / Code Mode for MCP — turn any OpenAPI spec into two sandboxed tools (search + execute).

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages