How do I connect Claude Code to Safari?

Update to Safari 27.0 (September 2026) on macOS, open Safari > Settings > Developer and tick 'Allow remote automation and external agents' (if there is no Developer tab, first enable 'Show features for web developers' under Advanced). Then run `claude mcp add safari-mcp -- "/usr/bin/safaridriver" --mcp` in your terminal. Start a Claude Code session and ask it to open your local dev URL in Safari; it can now take screenshots, read the DOM and console, inspect network requests and check accessibility.

How do I connect Codex to Safari MCP?

Same prerequisites — Safari 27.0 and the 'Allow remote automation and external agents' toggle — then run `codex mcp add safari-mcp -- "/usr/bin/safaridriver" --mcp`. For other agents, add a mcp.json or config.json entry: {"mcpServers": {"safari-mcp": {"command": "/usr/bin/safaridriver", "args": ["--mcp"]}}}.

Is Safari MCP safe to use?

It is the most conservative browser MCP available in 2026. The server runs entirely on your Mac, makes no network calls of its own, has no access to your personal information in Safari, and sends captured data to the agent you are running rather than to Apple. It uses the same isolated automation windows as Safari WebDriver, which start clean and cannot see your history or logged-in sessions. The risk that remains is the generic one: an agent driving a browser can be prompt-injected by page content, so point it only at sites you control or trust.

What can a coding agent do through Safari MCP?

See how code renders in Safari and capture screenshots; read and query the DOM; list network requests; read console output; verify user state in forms, checkout flows and selections; compare computed styles and layout against other browsers; test accessibility issues such as missing labels, improper ARIA and poor contrast; and analyse performance via navigation timing and resource load times.

Quick Answer

How to Connect Claude Code or Codex to Safari MCP (2026)

Published: September 19, 2026

What you get

Safari 27.0, released in September 2026, includes a built-in MCP server that lets a coding agent control a Safari window. Apple’s own list of what the agent can then do: see how your code renders, verify user state in forms and checkout flows, compare computed styles and layout with other browsers, test for accessibility issues, and analyse performance with navigation timing and resource load times. In practice it removes the loop of “run it, screenshot it, paste the screenshot into the terminal, describe what is wrong.”

The server is /usr/bin/safaridriver — the same binary Safari has used for WebDriver since Safari 10 — started with a new --mcp flag. It runs locally, makes no network calls of its own and does not touch your personal Safari data.

Prerequisites

macOS with Safari 27.0 or later. Check Safari > About Safari. (Safari 27 ships with macOS 27 and as an update for supported earlier macOS releases; if you do not see 27.0, run Software Update.)
Claude Code, Codex, or any MCP-capable agent (Cursor, Antigravity, OpenCode and others read a mcp.json).
A local dev server for the agent to load — e.g. http://localhost:3000.

Step 1 — Enable external agents in Safari

Open Safari > Settings.
If there is no Developer tab, go to Advanced and tick “Show features for web developers.”
Open the Developer tab and tick “Allow remote automation and external agents.”

Without this toggle safaridriver --mcp will start but refuse to open windows.

Step 2 — Register the server with your agent

Claude Code:

claude mcp add safari-mcp -- "/usr/bin/safaridriver" --mcp

Codex:

codex mcp add safari-mcp -- "/usr/bin/safaridriver" --mcp

Any other MCP client — add this to the client’s mcp.json or config.json:

{
  "mcpServers": {
    "safari-mcp": {
      "command": "/usr/bin/safaridriver",
      "args": ["--mcp"]
    }
  }
}

Verify with claude mcp list (or codex mcp list); safari-mcp should show as connected.

Step 3 — Give the agent a task that needs eyes

Start a session in your project and ask for something that requires the browser:

“Open http://localhost:3000/checkout in Safari, fill the form with test data, submit, and tell me whether the success state renders. Check the console for errors and list any failed network requests.”

The first time, Safari opens an automation window — marked with an orange Smart Search field — that starts from a clean slate with no history or logins. A transparent glass pane sits over it so stray clicks do not interfere. Let the agent work; it will read the DOM, capture screenshots and report back.

Useful follow-ups:

“Compare the computed styles of .hero h1 here and in Chrome and explain the difference.” (Pair with Chrome DevTools MCP.)
“Run an accessibility pass on this page: missing labels, ARIA misuse, contrast.”
“Report navigation timing and the five slowest resources on load.”

Step 4 — Keep it safe

Only point it at sites you control or trust. A page can contain text designed to hijack an agent that is reading it. Automation windows carry no logins, which limits the blast radius, but do not use the agent to browse arbitrary sites.
Untick the Developer toggle when you are done if the Mac is shared.
Do not enable every browser MCP at once. Each server adds dozens of tool definitions to every turn. Enable Safari MCP for WebKit checks; use Chrome DevTools MCP or Playwright MCP for profiling and headless test runs.
Keep your agent patched. Claude Code below 2.1.179 and Codex below 0.146.0 are exposed to the Plugin4Shell plugin vulnerability disclosed September 18, 2026.

Troubleshooting

Symptom	Fix
`safari-mcp` shows “failed” in `claude mcp list`	Confirm Safari is 27.0+, and the Developer toggle is on
Agent says it cannot open a window	Toggle “Allow remote automation and external agents” off and on; quit and relaunch Safari
Tools missing from the agent	Restart the agent session after adding the server; MCP servers are read at start-up
Page loads but screenshots are blank	Wait for load: ask the agent to check `document.readyState` before capturing
Works in Chrome, not Safari	That is the point — you have found a WebKit-specific bug; ask the agent to diff computed styles

Why this matters

Until September 2026 the only ways to give an agent a real Safari were third-party WebDriver wrappers or Playwright’s WebKit build, which is not Safari. Apple shipping a first-party MCP server means iOS-facing web teams can put “verify in Safari” inside the agent loop instead of at the end of it — and it sets an expectation that browsers will expose themselves to agents natively rather than through extensions.