Code Mode MCP
Composes arbitrary MCP tools using JavaScript, enabling discovery, composition, and execution of other MCP servers through a single 'exec' tool.
README
Code Mode MCP
Compose arbitrary MCP tools with JavaScript through one agent-agnostic stdio MCP server.
The server exposes one tool, exec. Upstream schemas stay out of the model's initial context: JavaScript finds relevant tools with ranked search(), inspects exact schemas with describe(), and invokes normalized functions on tools.
Independent and experimental. This is not an OpenAI or Pi product. The Code Mode API may change before version 1.0.
Upgrading from
pi-code-mode-mcp? SeeMIGRATION.md.
What it does
MCP client (Pi, Claude, Codex, or another host)
└─ exec({ code })
└─ standalone code-mode-mcp process
├─ tools.mcp__github__search_issues(...)
├─ tools.mcp__computer_use__get_app_state(...)
└─ Promise.all(...)
- one model-facing MCP tool instead of every upstream schema;
- stdio, Streamable HTTP, and legacy SSE upstream transports;
- JavaScript loops, branching, parallel calls, transformation, and filtering;
- in-code discovery and exact JSON Schema inspection;
- text, image, audio, resource, structured-content, error, and metadata forwarding;
- nested cancellation, progress, elicitation, sampling, roots, logging, and
tools/list_changedhandling; - bearer auth, OAuth client credentials, and interactive authorization-code OAuth;
- explicit, JSON-only, in-memory session state;
- no automatic persistence of tool results, screenshots, logs, or intermediate values.
Normal client tools remain available directly. Code Mode composes the MCP servers configured behind it; it does not replace the host's direct tools or convert host-native tools into nested MCP functions.
Requirements
- Node.js 22 or newer
- an MCP client that can launch a stdio server
Install
Install the exact npm release globally:
npm install --global code-mode-mcp@0.2.0
code-mode-mcp --help
Or build a source checkout:
git clone https://github.com/tmustier/code-mode-mcp.git
cd code-mode-mcp
npm ci
npm run prepublishOnly
The source executable is dist/cli.js after the build.
Configure upstream MCP servers
Create ~/.config/code-mode-mcp/mcp.json:
{
"settings": {
"executionTimeoutMs": 120000,
"requestTimeoutMs": 120000
},
"mcpServers": {
"computer-use": {
"command": "node",
"args": [
"/absolute/path/to/codex-computer-use-mcp/dist/mcp-server.js"
],
"requestTimeoutMs": 180000
},
"remote": {
"url": "https://example.com/mcp",
"auth": "bearer",
"bearerTokenEnv": "EXAMPLE_MCP_TOKEN"
}
}
}
Validate without starting MCP:
code-mode-mcp --check-config \
--config ~/.config/code-mode-mcp/mcp.json
The JSON summary excludes commands, arguments, headers, tokens, and environment values.
Configuration lookup
When --config is omitted, the first existing file wins:
$CODE_MODE_MCP_CONFIG./.code-mode-mcp.json~/.config/code-mode-mcp/mcp.json- legacy
~/.config/pi-code-mode-mcp/mcp.json
PI_CODE_MODE_MCP_CONFIG and PI_CODE_MODE_MCP_HOME remain compatibility fallbacks.
The file uses the standard mcpServers object. Each server defines exactly one of:
command, with optionalargs,env, andcwd;url, with optionaltransport,headers, and auth.
URL transport defaults to Streamable HTTP with SSE fallback. Set transport to "streamable-http" or "sse" to require one.
Strings support ${VAR} and exact $env:VAR environment expansion. Relative cwd and settings.stateDir paths resolve from the config file.
OAuth
{
"mcpServers": {
"linear": {
"url": "https://mcp.example.com/mcp",
"auth": "oauth",
"oauth": {
"grantType": "authorization_code",
"scope": "read write"
}
}
}
}
For authorization-code OAuth, the server opens a loopback callback and forwards the authorization URL through MCP URL elicitation. The outer client decides whether to open it. Unsupported interaction returns cancel; the server never invents accept or decline.
OAuth tokens, dynamic client registration, PKCE verifier, and discovery metadata are stored as mode-0600 files under settings.stateDir (default ~/.config/code-mode-mcp). No tool result is stored there.
Add to any MCP client
Configure the outer server in any client that can launch stdio MCP processes:
{
"mcpServers": {
"code-mode": {
"command": "npx",
"args": [
"-y",
"code-mode-mcp@0.2.0",
"--config",
"/Users/you/.config/code-mode-mcp/mcp.json"
]
}
}
}
Keep the upstream file separate. Do not configure Code Mode as its own upstream server.
Pi through pi-mcp-adapter
Pi can add lifecycle and direct-tool settings in ~/.pi/agent/mcp.json or .pi/mcp.json:
{
"mcpServers": {
"code-mode": {
"command": "npx",
"args": [
"-y",
"code-mode-mcp@0.2.0",
"--config",
"/Users/you/.config/code-mode-mcp/mcp.json"
],
"lifecycle": "lazy",
"requestTimeoutMs": 180000,
"directTools": ["exec"]
}
}
}
Restart or reload Pi after changing MCP configuration. The native Computer Use Pi extension and all normal Pi tools remain active alongside Code Mode.
exec API
Input:
{
"code": "return search('app screenshot accessibility', { limit: 5 });",
"session_id": "optional-session",
"timeout_ms": 120000,
"max_output_chars": 51200
}
code is a raw JavaScript async function body, not JSON-encoded source or a markdown fence.
Discover
return search("app screenshot accessibility", { limit: 5 });
search() ranks tool names and descriptions and returns compact { name, server, tool, title?, description, score } matches. Use a short keyword query; rephrase or remove a term if it returns no result. It accepts optional { server, limit } filters; the maximum limit is 50. Inspect one exact schema:
return describe("mcp__computer_use__get_app_state");
ALL_TOOLS remains a frozen complete inventory for deterministic enumeration or custom filtering when ranked search is insufficient.
ALL_SERVERS reports connection status and bounded error messages for enabled upstreams.
Compose
const apps = await tools.mcp__computer_use__list_apps({});
const selected = ["Calculator", "TextEdit"];
const states = await Promise.all(
selected.map(app => tools.mcp__computer_use__get_app_state({ app }))
);
return states.map((state, index) => ({
app: selected[index],
text: state.content.find(block => block.type === "text")?.text.slice(0, 500)
}));
Use call(name, args) when a name is selected dynamically.
Return rich output
Returning a complete MCP CallToolResult preserves its blocks and fields:
return await tools.mcp__computer_use__get_app_state({ app: "Calculator" });
Select output explicitly when intermediate results are large:
const result = await tools.mcp__computer_use__get_app_state({ app: "Calculator" });
text("Current Calculator state");
image(result.content.find(block => block.type === "image"), "original");
Helpers:
text(value)emits a text block;image(dataUrlOrMcpImage, detail?)emits an image;emit(contentBlock)emits any valid MCP content block;console.log()and related methods are captured and returned, not written to MCP stdout.
Returned text is bounded in memory. The server never spills full output to disk. Filter and aggregate inside the code cell for the best context efficiency.
Session state
store("cursor", { page: 2 });
return load("cursor");
store, load, and clearStore use explicit JSON-only, process-memory state. It disappears when the Code Mode server exits. The default session_id is "default".
Host authority and fault containment
Generated code intentionally has the same authority as this Node process. It can use process, require(), dynamic import(), fetch(), filesystem, network, environment, and child-process APIs. node:vm supplies a fresh context, captured console, tracked standard timers, and synchronous timeout interruption; it is not a security sandbox.
The standalone stdio process is the fault boundary. A synchronous tight loop is interrupted by node:vm. A loop that wedges the process after an asynchronous continuation may require the outer MCP client to terminate and restart the stdio server. This is why Code Mode is separate from Pi rather than an in-process extension.
See ARCHITECTURE.md, SECURITY.md, and ADR 0001.
Development
npm ci
npm run check
npm test
npm run prepublishOnly
npm pack --dry-run
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.