Runtime's MCP server gives an agent every Runtime product as tools. An agent connects with one browser approval and no key to copy. Claude Code, Codex, Cursor, ChatGPT, VS Code and any MCP client can create sandboxes, run code, read files and fork machines through the server (MCP).
Why it matters for AI agents
Before MCP, each AI application needed its own integration for each tool. MCP gives both sides one protocol. A tool is written once as an MCP server, and every MCP client can use it. The project describes MCP as "a USB-C port for AI applications". Anthropic open-sourced MCP on 25 November 2024. It is now supported by Claude, ChatGPT, Visual Studio Code, Cursor and many other clients.
MCP in facts
| Fact | Value |
|---|---|
| What it is | An open protocol between LLM applications and external data and tools |
| Message format | JSON-RPC 2.0, UTF-8 encoded |
| Roles | Hosts (LLM applications), clients (connectors inside a host), servers |
| Server features | Resources (context and data), prompts (templated messages), tools (functions) |
| Client features | Elicitation: a server asks the user for more information |
| Transports | stdio, to a client-launched subprocess; Streamable HTTP, to one endpoint |
| Latest revision | 2026-07-28 |
Source: the MCP specification at modelcontextprotocol.io, read on 25 October 2026. The specification treats tools as arbitrary code execution and requires hosts to obtain the user's consent before invoking one.
Runtime's MCP server
Add the server to Claude Code or Codex with one command:
Terminalclaude mcp add --scope user runtime -- npx -y withruntime mcpcodex mcp add runtime -- npx -y withruntime mcpThe first time, the server offers one tool, runtime_connect. The agent calls
it and shows you a link and a code. You approve Connect agent in your
browser. Every Runtime tool then appears, and you do not need to restart
anything. The command is a bridge. It forwards MCP messages to the remote server
over the saved connection, so no key is stored in the agent's configuration.
Clients that speak MCP over HTTP connect directly to
https://api.withruntime.com/mcp over Streamable HTTP. They sign in through
the browser with OAuth 2.1 and PKCE (remote connection).
| Tool | What it does |
|---|---|
runtime_sandboxes_create |
Create a sandbox and wait until it is ready |
runtime_sandboxes_exec |
Run a command and get its exit code and output |
runtime_sandboxes_files_write |
Write a file into the sandbox |
runtime_sandboxes_interpreter_run |
A notebook-style session in Python, JavaScript, TypeScript and more |
runtime_sandboxes_fork |
Copies of a sandbox as it is now, with its memory and processes |
runtime_sandboxes_network |
Read or replace a sandbox's network rules |
runtime_account |
The account and credit; whether this key is read-only, and its daily spending limit |
Tool names follow the pattern runtime_<product>_<verb>. On first contact, the
server sends instructions. They cover the quick start, safe retries and how to
report problems. The full list is in tools.
A sandbox can also run its own MCP servers, such as GitHub, Postgres or a Playwright browser. Your agent connects to them at URLs (MCP servers in a sandbox).
Related
Sources
Checked 25 October 2026.
- What is the Model Context Protocol?
- MCP specification
- MCP transports
- Anthropic, Introducing the Model Context Protocol, 25 November 2024
Facts on this page were checked on 25 October 2026.