Rill Browser

• Reading a library's release notes and pressing ⌘E to have Claude Code try the new feature in the project that uses it
• Pointing at a table row or chart method in an article and asking Codex to apply it to your own data while you keep reading
• Sending a page whose design you like to your portfolio project as a reference for a restyle
• Comparing up to six documentation or product tabs in one table that quotes each source
• Asking the agent to find flight options on the web, answering its questions in-page and finishing the payment yourself
• Searching your history by describing a page you half remember

Rill's bet is that the browser, not the terminal or an editor, is where many coding-agent tasks begin: a release note, an issue, a design you like. Building the hand-off into the browser, with automatic project routing and background runs, removes the copy-paste-and-switch step, and because it reuses your own Claude Code or Codex login, trying it costs nothing beyond what you already pay. The privacy page is unusually specific about what each feature sends and where files live, and its Limits section is candid that the purchase and posting guardrail is an instruction and that prompt injection is only reduced. The price of entry is switching your daily browser to a 0.1.x beta that only runs on Apple silicon Macs and has no documentation beyond its website. Developers comfortable with that will find the workflow distinctive; anyone handling sensitive accounts should keep agent web tasks on a short leash.

About Rill Browser

Rill, made by Manifold AI Inc., is a native Mac browser with no model or agent of its own: its AI features run the Claude Code or Codex CLI installed on your Mac, signed in to your own account, so there are no API keys and no new subscription. The core move is ⌘E on any page. You describe a task in plain words (try this method on my data, use this page's look for my portfolio, test this release note), and Rill sends it with the page or your selection to the right local project. It builds that project list from your Claude Code and Codex history, explains why it chose a project, and asks when it is not sure. The agent then runs in the background while you keep browsing: an At work tray shows progress with tool steps and diffs, questions appear as in-page bubbles you can answer, and you are told when the work is done. New tasks are read-only until you allow edits, and in git repositories each edit can be undone. Project agents that need the web read it in background tabs of their own, with a separate temporary cookie store and without your logins, and click or type there only when you allow it. Requests that are not code (book a flight, find something on a site) run as tasks on the web, in ordinary tabs with your sign-ins; the agent is instructed to stop before the step that pays, orders, sends, posts or deletes, and it cannot type into password, card or one-time-code fields. Around this sit browser features that also use your provider: answers about a page shown beside it, comparison tables across up to six tabs, AI tab groups, a history written up as day episodes, search for a page you only half remember, a map of the projects your agents have touched, and a beta voice mode. The tradeoffs: it is a 0.1.x beta for Apple silicon Macs on macOS 14 or later only, the irreversible-step guardrail is an instruction to the agent rather than a lock, Rill's own privacy page says prompt injection is reduced but not solved, and most AI features send page text to Anthropic or OpenAI under your account. There is no documentation beyond the website, privacy page and press kit.