MCP for coding agents: servers, config files, and tokens

The Model Context Protocol is an open-source standard for connecting AI applications to external systems. A host, such as a coding agent, runs one client per server, and they exchange JSON-RPC 2.0 messages. A server offers tools the model can call, resources the application can read, and prompts a person picks. The model decides to call a tool, but the host sends the request, and the specification says hosts must obtain explicit user consent first. What MCP is in Rig covers the same ground from the product side.

The root key is mcpServers in most clients, servers in VS Code’s own format, mcp in OpenCode, and mcp_servers in Codex. MCP config file locations lists paths, keys, and transports client by client. The MCP config generator writes the exact JSON or TOML for each client and converts a snippet from one client’s shape to another’s.

With stdio, the client starts the server as a local process. With Streamable HTTP, it sends requests to a URL. A token in a config file is a credential in a file, so treat the file like a .env file: write a variable reference where the client supports one, keep personal tokens in user-scope files, and commit a project file only when every value in it is safe for anyone who can read the repository. Env files and secrets covers the habits.

1. Read the Model Context Protocol entry for the vocabulary.
2. Find your agent in MCP config file locations, then generate its entry with the MCP config generator.
3. If you use Rig, follow Connect AI editors (MCP) and the page for your client, such as Cursor, Claude, or Codex.
4. Decide how much an agent may do, using Project-scoped MCP and the MCP tools overview.

Rig runs a local MCP hub that speaks Streamable HTTP on 127.0.0.1, port 47823 by default, and asks for a bearer token. It offers dozens of scoped tools for projects, ports, logs, env profiles, and git. Access comes in three modes: read, control, and full. Any write waits for an Approve click in Rig. The Link action writes a rig entry into the config of Cursor, Claude Code, Codex, Windsurf, VS Code, GitHub Copilot CLI, OpenCode, and Pi. It has no writer for Claude Desktop. See MCP that isn’t a raw shell for the design, and the MCP hub for Cursor and Claude and Cursor and Claude MCP pages for the product view.

Client formats change, so if a client rejects a snippet, the vendor’s documentation is the authority. The generator writes configuration text only: it does not start a server or test a connection, and it covers stdio and Streamable HTTP, not legacy SSE servers. The protocol moves too. Clients on older revisions still use a handshake that the current revision removed.