MCP

Configure Model Context Protocol servers for Claude Code, Codex, and Cursor Composer workflows

Conductor sessions can use Model Context Protocol (MCP) servers where the selected agent host supports MCP. Claude Code and Codex sessions load their agent-specific MCP configuration inside Conductor. Cursor Composer MCP servers are configured in Cursor and apply when you open a Conductor workspace in Cursor.

Use MCP when an agent needs repeatable access to external tools, private data, or workflow actions such as documentation search, issue tracking, database queries, or internal APIs.

For a step-by-step setup, see Set up MCP servers.

How MCP works

MCP uses a client-server model. The agent host starts or connects to one MCP server per configured integration, discovers the server's available tools, and routes tool calls to that server during the conversation.

Official MCP docs describe three core server primitives:

Primitive	What it provides	Example use in coding work
Tools	Executable functions the agent can call	Search docs, create an issue, query a database
Resources	Data sources the agent can read as context	Read API docs, schema metadata, or project records
Prompts	Reusable prompt templates exposed by the server	Run a project-specific workflow

See the official MCP architecture overview for the protocol model.

Transport types

Transport	Use it when	Notes
`stdio`	The server runs as a local command, such as `npx`, `node`, `python`, or a local binary.	The MCP client launches the process and exchanges JSON-RPC messages over stdin and stdout.
Streamable HTTP	The server is hosted at a URL and supports the current MCP HTTP transport.	Remote servers can use bearer tokens, OAuth, API keys, and custom headers.
SSE	You already depend on an older server that has not moved to Streamable HTTP.	Claude Code still documents SSE, but marks it deprecated in favor of HTTP where available.

The official MCP transports specification defines stdio and Streamable HTTP as the standard transports.

Configuration scopes

Scope	Claude Code	Codex	Cursor Composer	Use it for
User	`~/.claude.json` or `claude mcp add --scope user`	`~/.codex/config.toml` or `codex mcp add`	`~/.cursor/mcp.json` or Cursor's MCP settings UI	Personal MCP servers you want available across repositories.
Project	`.mcp.json` in the repository root	`.codex/config.toml` scoped to the trusted project	`.cursor/mcp.json` in the repository root	Shared project MCP servers that should follow the repository.
Local project	`~/.claude.json` under the current project path	Personal `~/.codex/config.toml` entries, optionally limited to trusted projects in the Codex config.	Local Cursor config that is not committed	Machine-specific credentials or experiments.

Conductor does not define a separate MCP config format. It uses the MCP configuration that Claude Code and Codex load for the session. Cursor Composer uses Cursor's MCP configuration when the workspace is open in Cursor.

Add a Claude Code MCP server

Add a user-scoped stdio server with the Claude Code CLI:

claude mcp add <server-name> -s user -- <command> [args...]

For example, add Context7 for documentation search:

claude mcp add context7 -s user -- npx -y @upstash/context7-mcp

Use the -- separator before the server command and arguments. Claude Code treats flags before -- as Claude options and everything after -- as the server command.

For more options, see the Claude Code MCP docs.

Add a Codex MCP server

Add a stdio server with the Codex CLI:

codex mcp add <server-name> -- <command> [args...]

For example, add Context7 for documentation search:

codex mcp add context7 -- npx -y @upstash/context7-mcp

Codex stores MCP configuration in ~/.codex/config.toml. The Codex CLI and IDE extension share that configuration.

You can also edit config.toml directly:

[mcp_servers.context7]
command = "npx"
args = ["-y", "@upstash/context7-mcp"]

Codex also supports streamable HTTP MCP servers:

codex mcp add openaiDeveloperDocs --url https://developers.openai.com/mcp

For more options, see the Codex MCP docs and Codex configuration reference.

Add a Cursor Composer MCP server

Add Cursor Composer MCP servers in Cursor's MCP settings UI or by editing mcp.json. Cursor supports user-level and project-level MCP configuration:

.cursor/mcp.json

{
    "mcpServers": {
        "context7": {
            "command": "npx",
            "args": ["-y", "@upstash/context7-mcp"]
        }
    }
}

Use .cursor/mcp.json when the MCP server should follow a repository. If you open a Conductor workspace in Cursor, Cursor reads the project-level .cursor/mcp.json from that workspace checkout.

For more options, see the Cursor MCP docs.

Project-level MCP servers

If a repository has an .mcp.json file at its project root, Claude Code sessions in that Conductor workspace inherit those MCP servers.

For Codex, you can scope MCP servers to a trusted project with .codex/config.toml.

For Cursor Composer, use .cursor/mcp.json at the repository root. This is Cursor's project-level MCP config file, not the same file as Claude Code's root .mcp.json.

If you're moving project-level MCP config from Cursor, see Work with Cursor and VS Code.

Status in Conductor

Conductor can show MCP server status for Claude Code and Codex in the chat composer and the MCP status dialog. Before a session starts, the status reflects the workspace-level configuration Conductor can discover. After a session starts, the status reflects the running session.

Use Refresh status after you change MCP configuration. For Codex, Conductor opens codex mcp list in a terminal when you run the MCP command flow, because the Codex CLI expects a concrete mcp subcommand.

Cursor Composer MCP server status is managed in Cursor's MCP UI. Conductor-hosted Cursor sessions do not currently surface Cursor MCP server status in Conductor's MCP status UI.

Good candidates

MCP works best for tools an agent needs repeatedly, such as documentation search, issue tracking, databases, and internal APIs.

Common examples include Context7 for documentation search and Linear for issue management.

Security and privacy

MCP servers can run local commands, read local data, and send data to external services depending on the server you configure. Review the server command, requested credentials, and tool list before adding it.

For local MCP servers, prefer trusted packages or local code you control. The official MCP security best practices recommend clear consent before running local server commands and warn that local MCP servers run with the user's privileges unless sandboxed.

Enterprise data privacy disables custom MCP servers. See Security and permissions.

On this page