The Pulse
Safari 27 Gives AI Coding Agents Direct Browser Control
Safari 27 introduces a local Model Context Protocol server that lets Claude Code, Codex and other agents inspect and test webpages inside Safari. The server exposes DOM data, network requests, screenshots and console output without sending
AI.info Team ·
Safari 27.0, released on September 17, adds a local Model Context Protocol server that gives AI coding agents direct access to a Safari browser window. WebKit says the feature lets Claude Code, Codex and other MCP-compatible agents inspect how webpages render instead of relying on screenshots or written descriptions from a developer.
The server exposes the browser’s document object model, network requests, screenshots and console output. Agents can use that information to check form states, compare computed styles, investigate layout differences, identify accessibility problems and examine navigation timing and resource-load performance. The feature is aimed at web development and debugging rather than general-purpose browser automation.
WebKit’s announcement is available in its Safari 27 feature report, published two days after the earlier release of Safari 27. The release notes describe the MCP server as one of the version’s headline additions among a much larger set of WebKit changes.
Safari Gives Agents More Than Screenshots
Browser-connected agents have generally depended on indirect evidence: a developer can paste an error, attach an image or describe what appears on screen. Safari’s MCP server gives the agent structured browser data as well as visual output, allowing it to inspect the page and its execution environment directly.
WebKit lists several uses for the connection. An agent can verify user states in forms and checkout flows, inspect selections, compare Safari’s computed styles with another browser, and look for missing labels, incorrect ARIA attributes or poor color contrast. It can also examine network activity and performance timings while a page loads.
The arrangement is deliberately local. WebKit says the server runs entirely on the user’s machine and makes no network calls of its own. It does not access personal information in Safari, and captured data goes directly to the agent the user has chosen rather than to Apple.
One Switch, Then an MCP Configuration
Developers must first enable Safari’s external-agent access under Safari > Settings > Developer > Allow remote automation and external agents. If the Developer section is hidden, Safari’s Advanced settings provide a separate option to show web-developer features.
WebKit provides command-line setup examples for both Claude Code and Codex. The commands register Safari’s WebDriver executable as an MCP server:
claude mcp add safari-mcp -- "/usr/bin/safaridriver" --mcp
codex mcp add safari-mcp -- "/usr/bin/safaridriver" --mcp
Other clients can use the same executable in an MCP or agent configuration file, with --mcp passed as its argument. The earlier WebKit introduction to Safari MCP lists browser actions including taking screenshots, changing the viewport size, emulating a CSS media type, switching tabs and waiting for navigation to finish.
A Browser Tool Inside a Larger Safari Release
Safari 27 includes more than the agent connection. WebKit says the release grew to 83 features from 58 in the first beta, while the number of resolved issues reached 844 after additional fixes were added to the 525 announced at WWDC.
The release also brings customizable native select controls, wider availability for the HTML <model> element, scroll anchoring, WebAssembly JavaScript Promise Integration and a rewritten ECMAScript module loader with support for top-level await. Web Inspector gains inline color-contrast information and a fuller view of redirect chains in the Network tab.
Safari 27.0 arrives with macOS 27, iOS 27, iPadOS 27 and visionOS 27. Safari 27.0 is also available as a separate update for macOS 26 Tahoe and macOS 15 Sequoia, according to WebKit’s release post.
Why the MCP Connection Matters to Web Developers
The practical change is not that an agent can click through a webpage. WebDriver has supported browser automation for years. Safari MCP combines that control with developer-facing inspection data, giving an agent a way to connect a visible rendering problem to the DOM, console, network and performance evidence behind it.
That makes Safari a more direct test environment for coding agents. A developer can ask an agent to investigate a broken layout or an inaccessible form while the agent gathers the relevant browser state itself, rather than requiring a sequence of manually captured artifacts.
WebKit’s implementation also keeps the boundary clear: Safari supplies browser data to the local agent, while the agent remains responsible for whatever model or service processes that information. For developers testing Safari-specific behavior, the new server is available through the same safaridriver tool already used in Apple’s browser automation stack.