One package · your choice of client

Install it.
Put it to work.

Choose the AI you use. Local coding clients connect to the same MCP server; ChatGPT has a separate connection path.

Local server: Node.js 20+. Simulator tools also require macOS and Xcode. Run setup from your app folder. Review tool permissions before allowing file changes or simulator actions.

Claude

In Claude Code, connect the local server:

claude mcp add ios-agent -- npx -y ios-agent-mcp@latest
  1. Restart or reconnect your Claude session and check that the MCP tools are available.
  2. Ask the agent to list the connected tools, search only the relevant local references, and review a small Swift file first. For changes, ask it to build and test, show the command results, and explain unresolved failures.

Claude Desktop: use the Desktop JSON setup instead of the CLI command. Preserve existing settings.

Codex

Connect the server to your local Codex coding environment:

codex mcp add ios-agent -- npx -y ios-agent-mcp@latest
  1. Open your app folder in Codex and reconnect the MCP server if needed.
  2. Ask the agent to list the connected tools, search only the relevant local references, and review a small Swift file first. For changes, ask it to build and test, show the command results, and explain unresolved failures.

Use the local coding environment for source edits and Xcode workflows. The ChatGPT web setup below is different.

ChatGPT

ChatGPT web cannot run the local npx command on your Mac directly.

  1. For guidance, use the skills-only release package through a supported import or plugin-development flow available to your account.
  2. For connected tools, configure an HTTPS knowledge MCP endpoint or a supported private MCP tunnel. This project does not operate a public endpoint for you.
  3. Follow the official connection guide (new tab). Account and workspace policy may limit access.
  4. Ask for a scoped implementation plan and relevant Apple references. Use an appropriately connected local coding environment for builds, simulator execution and evidence.

The hosted knowledge server serves references and plans. It is not a remotely hosted Mac or simulator.

Full ChatGPT setup ↗

Gemini CLI

Connect the published server from your project folder:

gemini mcp add ios-agent -- npx -y ios-agent-mcp@latest
  1. Restart Gemini CLI and confirm the server is connected.
  2. Ask the agent to list the connected tools, search only the relevant local references, and review a small Swift file first. For changes, ask it to build and test, show the command results, and explain unresolved failures.

This is for Gemini CLI, not Gemini web chat. The repository extension is an alternative for bundled guidance; check its version pin is published before installing it. Use one setup method to avoid duplicate tools.

Muse Code

Install the server outside the agent sandbox:

npm install -g ios-agent-mcp@latest

Merge this into ~/.config/muse/settings.json. Keep your other settings and existing servers.

{
  "schema_version": 1,
  "mcpServers": {
    "ios-agent": { "command": "ios-agent-mcp", "args": [] }
  }
}
  1. Restart Muse. If the executable is not found, use its absolute path from command -v ios-agent-mcp.
  2. Ask the agent to list the connected tools, search only the relevant local references, and review a small Swift file first. For changes, ask it to build and test, show the command results, and explain unresolved failures.

Muse tool discovery and a Stop hook have been verified; a complete model-driven app-building session has not.

Start small. Verify the result.

Try this prompt after your local MCP connection is working:

Review this app’s Swift concurrency. Search only the relevant local references.
Explain each finding with its file and line, and separate heuristics from confirmed bugs.
Propose the smallest fix. After approval, implement it, run the relevant tests,
and report what passed and what remains unverified.

When a step fails, ask the agent to show the relevant diagnostic, explain its next fix and report the result of verification. Keep this feedback in the development session, not in the app’s interface. Public issue reporting is separate: prepare a local preview, then review and authorize submission. A private chat-based option is being prepared in source; it is not deployed or available in npm 2.7.0.

For a new app, describe screens, persistence and acceptance criteria first. A starter is scaffolding; your agent still implements the features. Prefer one feature and one verification loop at a time.

Do not assume a connected tool means the app is complete. Ask for actual build/test results and simulator evidence where applicable.

Troubleshooting and full client documentation ↗