The browser where agents finish the work.

A full Chromium browser with an agent in the same window. Skylark works on the live tab, keeps the context that matters, and shows you every move.

skylark://start
Skylark on the desktop
  • A real Chromium browser The page is where the work happens
  • Agent on the live tab See, pause, and replay every action
  • Your model, your choice Bring a key, a local model, or your own agent
  • macOS build lands first No public download yet
What makes it different

Built for the moment an agent has to deliver

Most browser agents live outside the browser and lose the plot. Skylark starts with the page, the tab, and the controls that make real work possible.

  • Standards

    Open standards. Your way of working.

    Skills are plain text you can read, edit, and version. Sites can publish their own. Skylark speaks the standards that let your tools meet the browser, without locking your workflow into one vendor.

  • Memory

    A browser that remembers

    You remember the idea, not the URL. Skylark brings the page back, carries context through long tasks, and can turn a successful path into a reusable procedure.

  • Developer workflow

    Your coding agent, in the browser

    Use the CLI for a headless task or hand the live tab to Codex, Claude Code, Gemini CLI, or opencode. The browser becomes part of the workflow you already trust.

  • Trust

    See every move. Keep the last word.

    Every action is visible, replayable, and interruptible. Skylark asks before consequential steps, keeps credentials out of the agent’s view, and lets you draw a line before the work crosses it.

  • The agent

    It works from what is really there

    Skylark reads the page structure, follows links, fills forms, waits for change, and works from your open tabs. No pasted URLs. No guessing which page you mean.

  • Skills

    Know-how that travels with the site

    Sites can publish plain-text instructions and Skylark loads the relevant one when it matters. Teams can keep their own skills in Git, where the work is reviewable and reusable.

  • MCP

    A bridge, not a silo

    Skylark can expose the live browser to other MCP clients, or connect browser work to the tools you already run. The tab in front of you stays part of the system.

Memory

You remember the idea. Skylark finds the page.

Browser history records where you have been. Skylark helps you return to why it mattered, and keeps the task coherent while the work grows.

  • Recall

    Find the page from the thought you kept

    Describe the page you half-remember and Skylark brings it back. Page memory stays in a local database and uses the model you already configured, with no second model bill.

  • Long tasks

    Keep the thread when the work gets long

    Auto-compaction preserves the useful context as a task grows, so the agent can carry on instead of losing the plot halfway through.

  • Across sessions opt-in

    Pick up where you left off

    Turn on episodic memory and Skylark can recall what it did for you last Tuesday, across chats and across sites.

  • Learned procedures opt-in

    Teach it once. Reuse the path.

    Turn on learned procedures and a successful task can become a reusable skill, so the next run starts with a better map.

Page memory stays on your machine. Incognito leaves no durable record.

Reliability

An agent that keeps going until the page says it’s done

A browser agent earns trust by handling the ordinary ways browser work goes wrong: a changed page, an awkward form, a long chain of steps.

  • Planning

    Plan first. Adapt when the page changes.

    Skylark breaks a job into subgoals and constraints, then replans when the page does not behave as expected.

  • Verification

    Do not call it done without proof

    A self-verifier checks the result before the agent finishes. If the page has not proved the task is complete, the agent gets a specific reason to continue.

  • Recovery

    When one route stalls, try another

    Skylark can split work across fresh tabs and explore several strategies in parallel, then keep the approach that actually moves the task forward.

  • Page reading

    Read structure, not just pixels

    Skylark works from the DOM and accessibility tree: the labels, roles, and state that explain what the page can actually do.

  • Limits

    A budget for every run

    Each task works within limits for steps, time, and spend, so a confused agent cannot quietly run away with your attention or your tokens.

The agent

The page stays in view. So does the agent.

No copied URLs. No “which tab do you mean?” Skylark gives the agent the page in front of you and a place beside the work, so the conversation stays connected to the task.

  • Reads the page structure and your open tabs, with useful context.
  • Follows links, fills forms, scrolls, waits, and acts on the real page.
  • Every action appears live. Pause or interrupt at any point.
Read the architecture
skylark://start
Agent and page, side by side
Bring your own agent

Bring the agent you already trust

Skylark has an agent, and it can work with yours. Bring a coding agent into the live browser, or point an MCP client at the tab in front of you.

  • Bring your own

    Bring the agent you already trust

    Connect Codex, Claude Code, Gemini CLI, or opencode. Your coding agent can work through a real browser tab instead of an isolated browser abstraction.

  • MCP server

    Turn the browser into an MCP tool

    Connect an MCP client to Skylark and it can work with the live tab you are looking at. Use connected mode for the focused browser or headless mode for scripts and CI.

Benchmarks

We will earn the claim

Browser agents are easy to demo and hard to trust. We run the tasks that expose the difference, then publish the numbers when they are ready.

  • Live tasks

    Online-Mind2Web

    Live, multi-step web tasks.

  • Browser use

    BU-Bench-V1

    Browser-use scenarios.

  • Long horizon

    Odysseys

    Long-horizon planning.

No published score yet. We would rather earn the claim, then show the evidence.

Trust

Control belongs with you

Power is only useful when you can see it, steer it, and stop it. Skylark builds those choices into the work, not around it.

  • Visibility

    The work stays visible

    Every action is visible, replayable, and interruptible. You can see what happened, while it happens.

  • Credentials

    Credentials are not agent context

    Skylark can use your saved credentials without passing their values to the agent, and keeps credentials intended for one site from being used on another.

  • Permission

    Permission is not a blank cheque

    A skill can be offered without being enabled. Even when it can act, consequential steps such as a purchase, post, or deletion require your approval.

  • Boundaries

    Draw the line in plain language

    Set a boundary such as “stop before paying” and Skylark pauses for your call before it crosses that line.

How it works

From a tab to a finished task

The goal is simple: give an agent the real page, a clear outcome, and sensible limits.

  1. 01 Install

    Start with the browser you know

    When the macOS build lands, open it like any other browser. Your tabs, pages, and the web still behave like the web.

  2. 02 Connect

    Choose the intelligence behind it

    Add a key in skylark://settings/ai, use a local model, or connect an agent you already run. The choice is yours.

  3. 03 Ask

    Give it the outcome, not the clicks

    Ask in plain language: prepare for a meeting, catch up on a thread, complete the long form, or build the report.

  4. 04 Steer

    Watch it work. Step in whenever.

    Each action appears as it happens. Pause it, interrupt it, or set a stop-before boundary. You remain in control.

The work is on the page. Your agent should be too.

Skylark is early access for macOS. Put it to work on a task that matters.

Download

macOS comes first

There is no public build yet, on any platform. macOS comes first, then Linux and Windows. Watch the releases page and it will show up there before it shows up anywhere else.

All releases on GitHub