TabRunner

You give the goal. It runs the tabs. TabRunner is a browser agent that lives in your browser and works in it — not in a sandbox. It works the tab you're already on, uses your logged-in sessions, and reads, clicks and types on the sites you already use, until the task you described is done. Works in your real browser — your existing logins are its sessions. No setup on every site, no fake profile, no separate account. Bring your own provider — sign in with an AI subscription you already pay for, or paste an API key: 15 presets cover the major cloud and local providers, and a custom endpoint accepts anything speaking a standard API format. No vendor lock-in, no relay, no TabRunner server. Your credentials stay yours — a key or a sign-in goes straight from the extension to your provider. Nothing is stored outside Chrome. No account, no telemetry. Trusted input — clicks and keystrokes go through the Chrome DevTools Protocol, so they are genuine trusted events, not synthetic dispatches sites can ignore. See the work — a live plan, current action, token spend and elapsed time while the agent runs. Every step is logged in the chat. It remembers how you work — your standing instructions and the facts it learns along the way load into every task, per site or globally. Review, edit or delete all of it in Settings, or turn memory off entirely. Skills are saved recipes — turn a finished chat into a reusable skill, write one by hand, or import one as markdown from a URL. A skill tied to a site is offered automatically when a task starts there, and every skill is yours to edit, export, disable or delete in Settings. Set it on a schedule — a task can run once at a time you pick, daily, or every few minutes within the hours you allow. Each schedule keeps its own chat, so a recurring run can see what it did last time. Turn a task into a walkthrough — ask "and document it" and the finished task becomes a shareable, step-by-step guide with a screenshot of each step. Those screenshots stay on this computer and never enter the chat sent to the model, so a recording cannot reach your provider. If one comes out incomplete, the guide says so in its own introduction. On by default; switch it off in Settings. Two-way MCP — TabRunner speaks MCP in both directions. An AI client on your machine can hand it tasks through the local bridge, and TabRunner connects out to remote MCP servers you add: their tools join every task behind plan approval, and when a server needs your input it asks right in the panel. Run events — started, finished, errors, questions — can also POST to your own webhook URL, so tasks plug into your automations. HOW IT WORKS 1. Describe a task in the side panel — e.g. "pull the invoice from my inbox into the expense report". 2. Approve the plan it comes back with. TabRunner then works the tab like you would — reading the page, clicking, typing, scrolling — with real user input. 3. Watch it work step by step, or send it to the background and get a notification when it's done. PRIVATE BY DESIGN No TabRunner server exists. The extension speaks to your provider directly. Provider configs and chat history live in chrome.storage on this device. The model never receives raw HTML — it works from a compact semantic tree of the page, and sensitive fields (passwords, one-time codes, card numbers) never leave the page. Works on Chrome and other Chromium-based browsers. GUARDRAILS You approve the plan — TabRunner cannot click, type or navigate until you've okayed what it intends to do. Ask before acting — consequential actions (paying, sending, deleting) come back for your confirmation in the panel before they happen. Stop is real — Esc or the Stop button ends the task on the spot; anything you've already typed becomes the next task. Reasoning effort — pin none to max per task, or leave Auto and TabRunner runs the newest model your endpoint lists. LANGUAGES English · Português (Brasil) · Español. Light and dark theme, or follow your OS.