A real browser for your local AI agent.

Why

I wanted agents to test web apps via the browser. Everything I tried (Playwright, Puppeteer, anything else that wraps CDP) was too heavy and too unstable. CDP drops sessions. Playwright crashes on long runs. Chrome gets fatter every release. None of that is the actual problem though. CDP and Chrome were designed for humans staring at DevTools. They were never designed for an agent stuck in a while loop.

An agent pays per token. It blocks per response. It can't deal with the event firehose, and a 4kb DOM dump on every read burns the context budget fast. The Hacker News front page through Playwright is about 2000 input tokens before the agent has done anything. Through vibesurfer it's around 50.

vibesurfer is a native browser daemon in Rust. Reads return state tokens and tree deltas instead of the full DOM. Writes check the token. If anything moved between the read and the write, the call fails and the agent re-reads instead of clicking on a stale page. There are three real engines underneath: WKWebView on macOS, WebKitGTK on Linux, WebView2 on Windows. The protocol on top is text and line-oriented.

When you actually need pixels there's vs capture for screenshots, vs viewport to switch between mobile and desktop layouts, and vs layout to get bounding boxes. But text comes first.

Install

Homebrew (macOS, Linux):

brew tap frane/tap && brew install vibesurfer

curl:

curl -sSL https://raw.githubusercontent.com/frane/vibesurfer/main/install.sh | sh

Cargo:

From source:

git clone https://github.com/frane/vibesurfer && cd vibesurfer
cargo install --path crates/vs-cli

Linux needs WebKitGTK 6. Windows needs the WebView2 runtime (already on Windows 11, available for Windows 10 from Microsoft).

Wire it into your agent

Two integration paths, and they're independent. You can install either or both:

• Skill: drop SKILL.md into the agent's skills directory. The agent reads it as context and calls the vs binary directly through whatever shell it has. Use this for any agent that runs Bash but doesn't speak MCP.
• MCP: register vs mcp as an MCP server. The agent calls vibesurfer primitives as MCP tools over JSON-RPC, no shell required. Use this for agents with native MCP support.

The auto-installer does both where supported. After vs is on your PATH:

It detects Claude Desktop, Claude Code, Cursor, Codex CLI, Gemini CLI, and OpenClaw, then writes the SKILL.md plus the MCP entry into each one. Agents that only support one of the two get only the relevant piece. Re-run after upgrading.

Doing it by hand

For the skill path, copy skills/vibesurfer/SKILL.md from the repo into the agent's skills directory. For Claude-family agents that's typically ~/.claude/skills/vibesurfer/SKILL.md.

For the MCP path, add this block to the agent's MCP config (claude_desktop_config.json, .cursor/mcp.json, etc.):

{
  "mcpServers": {
    "vibesurfer": {
      "command": "vs",
      "args": ["mcp"]
    }
  }
}

Codex uses TOML with the same shape under [mcp_servers.vibesurfer]. The JSON form also sits at plugin/.mcp.json if you would rather copy it from the repo.

Per-agent shortcuts

Claude Code marketplace installs both surfaces from one command:

/plugin install frane/vibesurfer

Resolves .claude-plugin/marketplace.json at the repo root and plugin/.claude-plugin/plugin.json.

Gemini extension wires the MCP server plus the GEMINI.md context file:

gemini extensions install https://github.com/frane/vibesurfer

Reads gemini-extension.json at the repo root.

Not detected as automated

Modern anti-bot systems (Cloudflare, hCaptcha, reCAPTCHA, DataDome) gate first on event.isTrusted, then on movement timing, then on TLS/HTTP fingerprinting. vibesurfer's input dispatch is built to pass all three.

On macOS, vs act click and the coordinate primitives (vs click-at, vs hover-at, vs move-to, vs drag) route through native NSEvent mouseDown / mouseUp / mouseMoved on WKWebView. Each event carries isTrusted = true in JS, same as a real cursor click. The Bezier-pathed lead-in dispatched before every click reproduces Fitts-law arrival timing (digraph-derived control points, optional overshoot) so the visible motion looks like a human reaching the target rather than a teleport.

Since v0.1.11 the coordinate primitives are native on Linux and Windows too: XTest over x11rb on WebKitGTK (X11 / Xwayland; pure Wayland falls back to ENGINE_UNSUPPORTED), SendMouseInput on a WebView2 composition controller on Windows. All three engines emit isTrusted = true for cursor-primitive clicks. Ref-based vs act click is trusted on macOS only — on Linux and Windows it still dispatches through injected JS (isTrusted = false); use the coordinate primitives there for fingerprint-sensitive sites.

The walker also honors ARIA role="..." (Radix UI, Headless UI, Reach UI, every custom-div-as-button pattern), plus a tabindex heuristic for focusable divs/spans without a role. Modern React UIs surface as actionable refs without coordinate workarounds.

Short forms

Every primitive has a one-to-three-letter alias. Long forms exist for documentation; agent invocations should use the short form to save tokens.

Frequent flags: --session= / -S, --full / -F, --since= / -s, --limit= / -n, --page= / -P, --json / -j. Inspect subcommands have one-or-two-letter aliases too (i co for inspect console, i n for network, i req for request, i e for eval, i s for storage, i scr for scripts, i src for script, i d for dom, i p for performance).

Both forms work everywhere. The integration tests assert that the wire request from a short form is byte-identical to the wire request from the long form.

Quickstart

$ vs so                                        # session-open
@0                                             # state token (16 hex chars; 0 means none yet)
s_019e08a7…                                    # session id

$ vs o https://example.com                     # open the URL
@0                                             # the open call doesn't carry a snapshot
p_019e08a7…                                    # page id

$ vs v p_019e08a7…                             # view (snapshot the a11y tree)
@44d01704049d6d31                              # state token
1 doc "Example Domain"                         # ref 1, document
  0 el ""                                      # nameless wrapper
    2 hd "Example Domain"                      # ref 2, heading
    3 p  "This domain is for use in…"          # ref 3, paragraph
    5 p  "Learn more"                          # ref 5, paragraph
      4 lnk "Learn more" click,focus           # ref 4, link, supported ops

A snapshot is a list of refs. Each ref is an integer that survives across snapshots, so the agent can act on ref 4 ten turns later without re-reading the whole page. The two-letter codes (hd, p, lnk, btn, tf, …) compress the role into a few bytes instead of an ARIA string. Labels are in quotes; the trailing tokens after a label list which vs act operations the element supports. About twenty role codes total, listed in docs/PROTOCOL.md.

$ vs a 4 click                                 # act: click ref 4
@<new-token>                                   # new token, page mutated
?nav                                           # warning: navigation occurred
… new tree …                                   # the act response carries deltas;
                                               # on navigation it re-baselines to a full tree

vs act is the only mutating primitive. It takes a ref and an operation (click, fill, scroll, key, submit, hover, focus) and requires the most recent state token. If the page mutated between read and write (a JS timer fired, a websocket pushed an update, anything), the call returns ! STALE_TOKEN and the agent re-reads. No silent stale clicks. After a successful act on the same page (no navigation), the response carries only the deltas (+ref for adds, -ref for removes, ~ref for attribute changes), so a click that adds one button costs ~20 bytes on the wire instead of the whole DOM.

$ vs st                                        # status
session  s_019e08a7…  pages=1
page     p_019e08a7…  url=https://www.iana.org/help/example-domains  token=…

Every primitive call writes one row to a SQLite audit log before it returns. vs status reads that log. So does vs log. Replay, debugging, and governance all collapse to SQL queries against ~/.vibesurfer/state.db. There is no separate event stream to subscribe to.

The daemon auto-spawns on first call. State, captures, and downloads live under ~/.vibesurfer/. The transport is an AF_UNIX socket on Unix (~/.vibesurfer/daemon.sock) and a Windows named pipe on Windows; either way, the CLI handles the difference.

29 wire primitives total — the 19 core primitives are specified in docs/PRIMITIVES.md; the later additions (vs_inspect, the four cursor primitives, prompt-input, and the pending queue) are documented in the bundled SKILL.md and the CHANGELOG. The full wire format with every sigil and edge case is in docs/PROTOCOL.md. The per-platform per-primitive verification matrix is in docs/REALITY_CHECK.md.

Configuration

Build from source

Requires Rust 1.85+. Platform-specific dependencies:

• macOS (15+): nothing extra, links against system WebKit.
• Linux: libwebkitgtk-6.0-dev, libgtk-4-dev, libsoup-3.0-dev.
• Windows: WebView2 SDK pulled by webview2-com at build time; the WebView2 Runtime is required at run time.

git clone https://github.com/frane/vibesurfer && cd vibesurfer
cargo build --release

Run the test suite:

cargo test --workspace --lib --bins        # fast unit tests
cargo test --workspace                     # adds integration tests (real engine)

For Linux engine tests on a non-Linux host, use the Docker container. WebKitGTK 6's sandbox needs unprivileged user namespaces; the CI Linux job relaxes the AppArmor restriction with one sysctl on the bare runner, while the Docker fallback needs --privileged to do the same:

docker build -f Dockerfile.linux-test -t vs-test-linux .
docker run --rm --privileged -v "$PWD":/work vs-test-linux

See docs/DEVELOPMENT.md for the longer walkthrough.

The demo gif at the top of this README is a real interactive Claude Code session driving vibesurfer. To capture a fresh one, run docs/demo/record-claude.sh:

brew install asciinema agg
docs/demo/record-claude.sh         # writes docs/demo-claude.gif

The script enforces a TTY guard, isolates the demo home, and locks Claude to Bash so the agent must use the real vs binary (no MCP fallback, no built-in file tools). Each render is non-deterministic, since model output varies. The cached gif is committed so cloners and CI don't re-render.

Contributing

Issues and pull requests welcome. Open an issue first for anything beyond a small fix so we can discuss the approach. The codebase uses agented for transactional file edits during development; agented's workspace state is local-only (.agented/state.db) and is not committed.

Acknowledgments

Built on:

• objc2, macOS WebKit FFI.
• webkit6, Linux engine bindings.
• webview2-com, Windows COM layer.
• interprocess, cross-platform IPC transport.
• tiny_http, integration test fixture server.

Protocol borrows from agented, an editor for AI agents.

License

Apache-2.0. See LICENSE.