Use the same MCP servers in Claude Code, Codex and Cursor

Claude Code, Codex and Cursor all run the same MCP servers, but each stores them in its own file and format. This guide maps the formats side by side, converts one server across all three, and adds up the context cost of the tested agent tools and MCP servers you plan to enable. Every format claim links the vendor's official docs, checked 2026-10-26.

Want a ready mcpServers block instead? Our MCP config generator builds one from tested servers with each server's env var names.

Config formats and locations

Scroll sideways for Codex and Cursor →

Claude Code scopes, from its docs: local (default; this project, only you, in ~/.claude.json), project (.mcp.json, shared through version control) and user (all your projects, in ~/.claude.json). If one name is defined in more than one scope, local wins over project, project over user, and entries are not merged (Anthropic: Claude Code MCP docs). Codex says the ChatGPT desktop app, Codex CLI and IDE extension share the same config.toml (OpenAI: Codex MCP docs). Cursor reads project .cursor/mcp.json and global ~/.cursor/mcp.json (Cursor: MCP docs).

stdio vs remote HTTP

A stdio server is a command the client starts on your machine (npx, uvx, a binary). A remote server is a URL using Streamable HTTP. All three clients support both (Anthropic: Claude Code MCP docs; OpenAI: Codex MCP docs; Cursor: MCP docs). The difference is how you mark a remote entry: Claude Code needs "type": "http" next to url, and skips an entry that has a url but no type (it prints a warning and the other servers still load); Codex treats a table with url as Streamable HTTP; Cursor accepts url on its own. SSE is deprecated in Claude Code's docs. More on running remote servers: how to host an MCP server.

Worked example: Context7 in all three clients

Context7 (@upstash/context7-mcp) is a server we test (listing) and the example Codex's own docs use. The stdio version needs no key. Add it the same way in each client:

Claude Code

CLI (local scope; add --scope project to write .mcp.json):

claude mcp add --transport stdio context7 -- npx -y @upstash/context7-mcp

Or in .mcp.json:

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

Codex

CLI:

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

Or in ~/.codex/config.toml:

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

Cursor

In .cursor/mcp.json or ~/.cursor/mcp.json:

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

Remote variant with an API key

Context7's README gives a remote URL, https://mcp.context7.com/mcp, with the key in an Authorization: Bearer header (Context7 README). Export CONTEXT7_API_KEY in your shell, then reference it instead of pasting it. We have tested the stdio package, not this remote endpoint.

Claude Code .mcp.json:

{
  "mcpServers": {
    "context7": {
      "type": "http",
      "url": "https://mcp.context7.com/mcp",
      "headers": { "Authorization": "Bearer ${CONTEXT7_API_KEY}" }
    }
  }
}

Codex config.toml:

[mcp_servers.context7]
url = "https://mcp.context7.com/mcp"
bearer_token_env_var = "CONTEXT7_API_KEY"

Cursor mcp.json:

{
  "mcpServers": {
    "context7": {
      "url": "https://mcp.context7.com/mcp",
      "headers": { "Authorization": "Bearer ${env:CONTEXT7_API_KEY}" }
    }
  }
}

Env vars and secrets

• Claude Code expands ${VAR} and ${VAR:-default} in command, args, env, url and headers. An unset variable with no default stays as literal text, with a warning. In a remote server's url and headers, Claude Code's own and cloud-provider credentials such as ANTHROPIC_API_KEY always read as empty; copy the value into a variable with your own name (Anthropic: Claude Code MCP docs).
• Codex has no ${} syntax in the docs we read: set values in [mcp_servers.name.env], list names to forward in env_vars, or name the variable in bearer_token_env_var / env_http_headers. A client secret passed to codex mcp add --oauth-client-secret is saved in config.toml, so keep that file private (OpenAI: Codex MCP docs).
• Cursor uses ${env:NAME} in command, args, env, url and headers; stdio servers can also load an envFile, remote servers cannot (Cursor: MCP docs).
• See our API key rotation checklist and MCP servers that need an API key or sign-in.

Gotchas

• Headless runs load project servers. Anthropic's docs say that in claude -p runs and Agent SDK sessions Claude Code cannot show the approval prompt and loads project-scoped servers from .mcp.json without asking. To keep one out, add it to disabledMcpjsonServers, exclude project settings with --setting-sources, or use --strict-mcp-config with --mcp-config (Anthropic: Claude Code MCP docs). Check this before running claude -p in a repo you cloned.
• A url without type is skipped in Claude Code. Claude Code prints a warning (Skipped — invalid MCP server config) and the other servers still load, so that server just doesn't appear. Add "type": "http" to Cursor-style remote entries.
• Different variable syntax. ${VAR} (Claude Code) vs ${env:VAR} (Cursor). Copying one into the other leaves a literal string.
• Codex project config needs a trusted project, per its docs.
• Server names: claude mcp accepts only letters, numbers, hyphens and underscores.
• Timeouts differ: Codex defaults to 10 s startup and 60 s per tool call (startup_timeout_sec, tool_timeout_sec); Claude Code uses MCP_TIMEOUT and a per-server timeout.
• -- matters in both claude mcp add and codex mcp add: everything after it is the server command.

Context budget picker

Every server you enable sends its tool definitions to the model. Tick the servers you plan to keep on in all three clients to see the sum. Figures are the same measured numbers as our lowest context cost ranking: 107 of 187 servers we list have a measured tools/list; the rest show not measured and add nothing.

FAQ

Can Claude Code, Codex and Cursor share one config file?

No. Claude Code and Cursor both use JSON with an mcpServers key, but in different files; Codex uses TOML in config.toml. The server command or URL is the same in all three, so you copy the entry and convert the format.

Where does each client keep project MCP servers?

Claude Code: .mcp.json in the project root. Codex: .codex/config.toml in the project, for trusted projects only. Cursor: .cursor/mcp.json in the project. Checked on each vendor's docs on Oct 26, 2026.

Does claude -p start project .mcp.json servers without asking?

Yes, per Anthropic's docs: in claude -p runs, Agent SDK sessions and some cloud sessions, Claude Code cannot show the approval prompt and loads project-scoped servers without asking. Block one with disabledMcpjsonServers, or start with --strict-mcp-config and pass only --mcp-config.

How do I keep API keys out of a shared config?

Reference an environment variable instead of pasting the value: ${VAR} in Claude Code's .mcp.json, ${env:VAR} in Cursor's mcp.json, and bearer_token_env_var, env_vars or env_http_headers in Codex. Do not commit files that hold real keys.

How many MCP servers should I turn on?

There is no single number. Use the context budget picker on this page to add up the tool definitions of the servers you want, then turn off what you do not use in that project.

No product here is rated or ranked, and this page has no affiliate links. How we test: methodology.

Sources (accessed 2026-10-26)

• Anthropic: Connect Claude Code to tools via MCP
• OpenAI: Codex Model Context Protocol
• Cursor: Model Context Protocol (MCP)
• Upstash Context7 README

Docs checked 2026-10-26