ClawdCursor

What's new

1.5.27 — Results you can trust, and logs that say what happened.

• Ticks, dropdowns and expanders read and change the same way on Windows, macOS and Linux; toggle reports the state it actually reached, and actions by name aim at the control, not its label.

• A stop always answers "stopped by the user" first, and also blocks hover, clipboard writes, page script and tab switches.

• Typing goes only to the window the agent was working in, dialogs like Save As included.

• Linux reads deep web pages, and a search that runs out of time or elements says so instead of "not found".

• Logs in ~/.clawdcursor/logs/ show every call's timing and outcome, slow internal steps, why the server exited, and who stopped it, with no typed text, error text or window titles.

1.5.26 — The agent asks you on screen, and the pill says what it is doing.

• Security (Windows, high): text the agent typed or copied could run as PowerShell code (GHSA-9ppf-38v8-h97v). Fixed; upgrade.

• The pause prompt is a card: the agent can offer 2-6 choices to click, or a text box, and gets your exact answer back. A choice is never a safety approval.

• A redesigned pill: what the agent is doing right now ("Typing in Edge"), how long it has been running, and a Stop button.

• A field below the fold is scrolled into view instead of refused (verified on Windows; the macOS and Linux calls are in but not yet run on a real desktop).

• macOS: when a password field holds secure input, window focus names the app holding it instead of failing vaguely.

1.5.25 — A stop really stops, and keys only go to the window the agent was working in.

• Double-click the pill and every clawdcursor process stops: a running wait ends at once, the rest of a batch never runs, and it stays stopped (even across restarts) until you click Resume.

• Every key, paste and click first checks that the agent's window is still in front. If you switched apps, nothing is sent and the agent is told so.

• Safety gate: key combos are blocked however their modifiers are spelled, shortcuts_execute and compound actions are gated like the tools they run, and every element an invoke may act on is checked.

• --no-vision now really keeps screenshots from every model; the daemon is loopback-only (use an SSH tunnel for remote), and server.port in config works.

• macOS: the Screen Recording check works without the helper built, and doctor asks the same question screenshots do.

1.5.24 — The AI never touches your pill, and failures stop looking like results.

• While a pause waits for you, nothing reads, clicks or types. The pill and pause prompt are never a target, even from another clawdcursor process.

• On macOS, a screenshot without Screen Recording permission is an error, not a wallpaper image, and open popovers are readable.

• The browser clawdcursor drives comes to the front when it navigates.

• A failed OCR read reports its error instead of "No text detected", and OCR confidence is only shown when the engine really has one.

1.5.23 — Windows: finding an element by name no longer matches stray letters. A lone "C" on another page could match "User code 0", and "Don't Save" could match Save. A partial match now has to be a real part of the name you asked for.

1.5.22 — Your turn is as short as possible. When an agent needs you (a sign-in, a 2FA code), it asks for that one step and does the rest itself. With done_when (what shows on screen once you're done), the prompt turns from blue back to amber the moment you finish, with no Resume click.

1.5.21

• set_value never guesses a field. With no field name or role it used to type into the window's largest text box; it now asks you to name the field.

• macOS: no screen-recording permission no longer breaks everything. Only screenshots and OCR fail, and they say how to fix it; windows, keys and accessibility keep working.

• Cleaner page text on Windows (no placeholder characters), and a security update to a dependency.

1.5.20

• The on-screen pill no longer blocks the agent's own clicks. It sits where browsers keep their tabs and address bar; it now steps aside for a moment before every click, drag or scroll, then comes back. Windows, macOS and Linux.

• Reading a page no longer highlights it on Linux and macOS either. copy_all_text reads the document through accessibility, as on Windows: exact text, nothing selected, your clipboard untouched.

• Your clipboard is put back after typing (smart_type too), and only once the paste has landed, so a slow app can't paste what you had copied.

• macOS: window focus focuses the process you asked for, a locked screen says so, sidebar clicks are verified, and a bare modifier key (shift) works.

1.5.19 — Typing never pastes your own clipboard. clawdcursor types by pasting; on Windows a clipboard write could fail silently and the paste then sent whatever you had copied. Every paste now checks the clipboard first and types key by key when it doesn't hold the text.

1.5.18

• "Something else" on the pause prompt. When the agent pauses for you, the prompt has a third button between Resume and End task: type an answer instead of doing the step — "use the test account instead", a code, a change of plan — and press Enter. The agent continues from what you wrote. Windows, macOS and Linux; over MCP pause_for_user returns {status:"replied", reply}.

• Reading a page no longer highlights it (Windows). copy_all_text reads the document the way a screen reader does: exact text, no Ctrl+A, nothing selected, your clipboard untouched.

• MCP tool annotations on every tool, and window / system now describe every action, its side effects and what it returns.

1.5.17

• macOS: text outside ASCII is typed and pasted as itself — "café 日本語" no longer comes out as "cafa aaa" or "caf√©".

• copy_all_text never selects an editor or a list (Ctrl+A there selected a whole document or every email).

1.5.16

• It pauses for you at a roadblock, then carries on. A sign-in, a 2FA code, a CAPTCHA, an upgrade page: instead of giving up, the agent pauses. The glow around the screen turns blue and an on-screen prompt says what it needs ("Sign in to Outlook, then click Resume"). Click Resume and it continues where it left off, or End task to stop. Your own agent can do the same over MCP with system({ action: "pause_for_user", message }); from a terminal, clawdcursor resume (or --end).

• The glow and banner on every OS. macOS and Linux now get the same screen-edge glow and banner as Windows: amber while the agent works, blue while it waits for you. Double-click the banner to stop everything.

• Blind mode on macOS caught up with Windows. A native accessibility helper lists windows in 0.3 s (was 11 s) and reads a window's tree in 0.06–0.2 s. Typing is refused while a system dialog owns the keyboard, and the agent is told when another app's dialog took its keys.

• Sharper blind mode on Windows and Linux. Classic controls no longer read as blank panes, the open dialog is what gets read, no 10 s stall on buttons that open a modal, the right field is found by name, and focusing a window no longer un-maximizes it.

• Upgrade walls are a roadblock. Clicks on Upgrade, Subscribe or Start trial need confirmation.

• Know what you're running. clawdcursor versions lists every installed copy (global npm, Claude Desktop extension, Claude Code plugin), and the installers say what they updated.

Everything else is in the changelog.

Related MCP server: windows-computer-use-mcp

What it is

Clawd Cursor is a local MCP server that gives any tool-calling agent — Claude Code, Cursor, Windsurf, OpenClaw, the Claude Agent SDK, or your own loop — safe control of the real desktop. It clicks, types, reads the screen, opens apps, and drives any GUI the way a human would: native apps, the browser, even a canvas.

Most "let an agent use the computer" tools take a screenshot and feed it to a vision model — slow, expensive, and brittle. Clawd Cursor compiles the screen into one UI map: it fuses the accessibility tree and OCR into a confidence-scored set of elements, each tagged with a stable el_NN id, and acts on elements by id — not pixel coordinates. Coordinates appear only in the last-resort screenshot/vision tier (live pixels off the current frame), for canvas-only apps or tasks that genuinely need spatial reasoning. The result is cheaper, faster, private, and — uniquely — it checks that each action actually did what it claimed.

If a human can do it on a screen, your agent can too. No API, no integration, no problem — only the right sequence of reads, clicks, keys, and waits. Use it as the last-mile fallback: native API exists? Use it. CLI? Use it. Clawd Cursor is for the click, the legacy app, the GUI with no public surface.

Why it's different

The desktop-agent space is crowded. The closest install-and-go peers are Windows-MCP and Terminator (desktop MCP servers); browser-only tools (browser-use, Playwright MCP) are adjacent; and OmniParser / UI-TARS are vision-centric parsing approaches you'd build an agent around, not products you install. Here's the honest comparison across those approaches — what Clawd Cursor does that the popular options don't:

Three things here are genuinely rare:

1. Cheapest-tier-first perception, fully local. Accessibility tree (free) → OCR (cheap) → screenshot (expensive — the only tier that puts pixels in the model's context; "screenshot" and "vision" are the same step). The agent climbs only when it must, so token cost tracks task difficulty — and with a local model, nothing leaves the machine. Vision-centric agents (OmniParser, UI-TARS) need a screenshot in the model for every observation.

2. It verifies. Pass expect on a consequential action and Clawd Cursor re-checks the live screen (with a short settle window for async UIs) and reports a DEVIATION instead of a hollow "success." A completed task can't be marked done on evidence that was already true before it acted.

3. One safety gate. Every call — from an editor over stdio, an external agent over HTTP, or the built-in loop — routes through a single safety.evaluate() chokepoint (allow / confirm / block) before it touches the desktop. The agent cannot bypass it.

Plus: whenever an agent is driving, an on-screen banner that says what it is doing ("Typing in Edge") and an amber glow around the screen edges, on Windows, macOS and Linux — press Stop on the banner to stop. A stop holds until you give control back: the agent's next request comes up as a prompt, and only your Resume lets it act again (or run clawdcursor resume). Keys and clicks also go only to the window the agent was working in: if you switch apps while it waits, its next input is refused instead of landing in your window. When the agent needs you (a sign-in, a 2FA code), the glow turns blue and the banner says what to do, with Resume and End task. A human at the machine always knows, and always has a kill switch.

Install

clawdcursor is an MCP server published to npm — install it into any MCP-capable agent (Claude Code, Claude Desktop, Cursor, Windsurf, Zed, OpenAI Codex, or your own loop) the same way you install any other MCP server.

Claude Desktop — one-click extension (no terminal)

Download the extension for your OS and open it — Claude Desktop installs it:

1. Double-click the file (or Settings → Extensions → Advanced settings → Install Extension… and pick it).

2. Turn on Allow clawdcursor to control this computer — that is the one-time desktop-control consent; until it is on, every tool returns a consent prompt.

3. macOS: the first time macOS asks, allow Claude under Privacy & Security → Accessibility and Screen Recording.

No Node.js, npm or JSON editing needed — Claude Desktop runs the bundled server itself. Everything below is for other agents, or for Claude Desktop without the extension.

1 — Install the engine + grant consent (once)

npm i -g clawdcursor
clawdcursor consent --accept    # one-time desktop-control consent (required)
clawdcursor grant               # macOS only — approve Accessibility + Screen Recording

Hosts that support MCP elicitation (e.g. Claude Code, VS Code) also let you skip consent --accept: the first tool call asks "Allow clawdcursor to control this computer" right in the app.

Zero-install also works — swap clawdcursor for npx -y clawdcursor in any snippet below and npx fetches it on demand. A global install is recommended anyway: it's pinnable and inspectable on disk (safer for a tool with full desktop control than auto-fetching latest every run), and it's the path on which the macOS native helper builds at install time. Requires Node.js 20+.

Per-OS prerequisites. Windows installs clean — sharp and @nut-tree-fork/nut-js ship prebuilt binaries, so no C++/Python build tools are needed. macOS needs Xcode Command Line Tools (xcode-select --install) for screenshots / vision; core accessibility-driven control still works without them. Linux needs a few system packages npm can't install: tesseract-ocr (OCR), python3-gi + gir1.2-atspi-2.0 (accessibility tree), and — on Wayland — ydotool (synthetic input).

2 — Add it to your agent (pick your host)

One click (Cursor, VS Code):

Every host runs the same command — clawdcursor mcp --compact, or npx -y clawdcursor mcp --compact with no global install. Only the config shape differs between hosts:

Claude Code

claude mcp add clawdcursor -s user -- clawdcursor mcp --compact

Cursor · Windsurf · Claude Desktop · Cline · Gemini CLI — the common mcpServers shape:

{
  "mcpServers": {
    "clawdcursor": { "command": "clawdcursor", "args": ["mcp", "--compact"] }
  }
}

Claude Desktop reads %APPDATA%\Claude\claude_desktop_config.json (Windows) or ~/Library/Application Support/Claude/claude_desktop_config.json (macOS). Gemini CLI reads ~/.gemini/settings.json.

VS Code (Copilot) — .vscode/mcp.json uses servers, not mcpServers:

{
  "servers": {
    "clawdcursor": { "command": "clawdcursor", "args": ["mcp", "--compact"] }
  }
}

Zed — context_servers in settings.json:

{
  "context_servers": {
    "clawdcursor": { "command": "clawdcursor", "args": ["mcp", "--compact"] }
  }
}

OpenAI Codex — ~/.codex/config.toml. The first launch can take a few seconds, so raise the 10 s default startup timeout:

[mcp_servers.clawdcursor]
command = "clawdcursor"
args = ["mcp", "--compact"]
startup_timeout_sec = 30

opencode — opencode.json uses mcp, and command is an array:

{
  "mcp": {
    "clawdcursor": { "type": "local", "command": ["clawdcursor", "mcp", "--compact"] }
  }
}

Any other MCP host (Goose, Continue, …) — register a stdio server with command clawdcursor and args mcp --compact.

Windows: some hosts start the server without a shell, and then can't find clawdcursor or npx (both are .cmd shims). If the server won't start, wrap it in cmd /c — e.g. "command": "cmd", "args": ["/c", "clawdcursor", "mcp", "--compact"], or claude mcp add clawdcursor -s user -- cmd /c clawdcursor mcp --compact.

That's the whole setup. Ask your agent: "open Outlook and reply to the latest email from Sarah."

Or: one-command plugin (Claude Code)

Skip the manual config — this repo ships a plugin that registers the tools and bundles the usage skill in one step. It resolves the package's bin (never a hard-coded dist/ path), so an upgrade can't break it:

claude plugin marketplace add AmrDab/clawdcursor
claude plugin install clawdcursor@clawdcursor

# Windows (PowerShell)
powershell -c "irm https://clawdcursor.com/install.ps1 | iex"

# macOS / Linux
curl -fsSL https://clawdcursor.com/install.sh | bash

Notes. You never run clawdcursor mcp yourself — the host spawns it over stdio on demand. clawdcursor doctor is not part of MCP setup; it only configures the built-in LLM for the autonomous agent daemon. On macOS, Accessibility is required (primary control path); Screen Recording is optional (vision fallback only). For editor permission allowlists, use the server-level wildcard mcp__clawdcursor rather than per-tool entries — it survives tool renames.

The engine

The perception + verification core (the UI State Compiler, since v1.5.0):

• compile_ui fuses the accessibility tree and OCR into one confidence-scored map of the screen, every element tagged with a stable el_NN id. Act on an element by {element_id, snapshot_id} instead of pixels — near-free in tokens, and it survives DPI, resize, and layout shifts. find_button / find_field locate a target by meaning and hand you the id.

• Reactive verification. expect on an action → Clawd Cursor confirms the outcome on the live screen and returns a DEVIATION when the UI didn't obey.

• Cross-platform parity. The compiler, secure-field redaction, and coordinate handling run on Windows, macOS, and Linux; the external-agent (MCP) surface resolves el_NN refs through the safety gate and discloses when it attached to your existing browser.

Set-of-Mark-style element IDs and a11y/OCR fusion aren't new ideas on their own — what's rare is doing them locally, a11y-first (no vision model required), with a built-in verification gate and one safety chokepoint, across three operating systems, behind a single MCP config.

See the changelog for the full release history, or the latest release.

How it works

Where the brain lives decides how you run it. Both modes can run side-by-side.

Deprecated in 1.5.27: the built-in LLM loop (task, submit_task, system delegate and the clawdcursor agent model set up by doctor) will be removed in 1.6, and clawdcursor will no longer call a model itself. Drive the MCP tools from your own agent instead (Claude Code, Codex, Cursor, …). The daemon's HTTP MCP transport stays for remote use.

The loop

Read the a11y tree (cheap) → act on named targets → verify from fresh observations → escalate perception only when needed (OCR → screenshot, the one tier that sends pixels to the model). Sparse a11y tree? system.detect_webview switches Electron/WebView2 apps to browser.* over CDP. Canvas-only (Paint, Figma, games)? Screenshot + coordinate click.

flowchart TB
    task["User task"] --> loop["Agent LLM loop<br/>plans · chooses tools · verifies"]
    loop --> observe{"Cheapest observation<br/>that answers the question"}

    observe -- "obs·a11y — free" --> a11y["A11y tree<br/>(structured text + el_NN handles)"]
    observe -- "obs·ocr — cheap" --> ocr["OCR (OS-level, no vision LLM)"]
    observe -- "obs·dom — medium" --> dom["Browser DOM (CDP)"]
    observe -- "obs·vision — expensive" --> vision["Screenshot (image into context)"]

    a11y --> act
    ocr --> act
    dom --> act
    vision --> act

    act["Act<br/>click/type/key/drag · invoke/set_value · open_app · batch"] --> safety
    safety["Single safety gate<br/>safety.evaluate() → allow / confirm / block"] -- allowed --> tools["Tool registry<br/>100+ granular + 7 compound"]
    safety -- needs user --> confirm["Human confirmation"] --> tools
    safety -- denied --> blocked["blocked"]

    tools --> desktop["Real desktop"]
    desktop --> verify{"expect → does state match?"}
    verify -- pass --> done["done"]
    verify -- "DEVIATION" --> loop

    classDef agentNode fill:#dbeafe,stroke:#2563eb,color:#0f172a;
    classDef gate fill:#ede9fe,stroke:#7c3aed,color:#0f172a;
    classDef obsNode fill:#fef9c3,stroke:#ca8a04,color:#0f172a;
    classDef actNode fill:#ffedd5,stroke:#ea580c,color:#0f172a;
    classDef stop fill:#fee2e2,stroke:#dc2626,color:#0f172a;
    class loop,verify agentNode;
    class safety,confirm,tools gate;
    class observe,a11y,ocr,dom,vision obsNode;
    class act actNode;
    class blocked stop;

batch for deterministic stretches. When the next N steps are known, collapse them into one call — each step still routes through the safety gate; on any guard miss or error the batch halts with a per-step trace.

Task delegation (deprecated, removed in 1.6). With an LLM configured on the daemon, an external agent can hand off at any point: task({"instruction":"…"}). The built-in loop takes the wheel and reports back — offload grunt work to a cheaper model without burning your own context.

The toolbox

Two catalogs ship side-by-side. The toolbox is 7 compound tools, each with an action enum covering ~10–20 verbs (about 7,000 tokens for the whole catalog, schemas included — about a third of the granular surface's ~20,000; the computer_20250124 shape editor hosts already know). The granular surface is the 100+ underlying primitives, one schema per verb (for runtimes that need top-level tools, or for debugging). Both run through the same safety.evaluate() chokepoint; the full catalog is always visible via MCP tools/list.

computer({ action: "key", combo: "mod+s" })          // Cmd+S / Ctrl+S, resolved per-OS
accessibility({ action: "invoke", name: "Send" })    // click by name, not pixels
window({ action: "open_app", name: "Outlook" })

Cheapest-tier-first perception

Every observation has a cost. Start at the cheapest rung that works; climb only when it fails. The live log (CLAWD_LOG=pretty, default on a TTY) shows the ladder in real time via per-call badges.

Acting tools log act. Watching obs·a11y → act → obs·a11y on a normal turn — and the rare climb to obs·vision — is the whole efficiency model, visible.

Transports

One protocol — MCP — two transports, same catalog and JSON-RPC envelope. Both stateless; no session handshake.

# HTTP MCP — list tools
# (the Accept header is required — the MCP spec's Streamable HTTP transport
#  rejects requests that don't accept both JSON and SSE with a 406)
curl -s -X POST http://127.0.0.1:3847/mcp \
  -H "Authorization: Bearer $(cat ~/.clawdcursor/token)" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Platform support

Platform code lives behind a single PlatformAdapter interface (src/platform/{windows,macos,linux}.ts + wayland-backend.ts). Business logic never reads process.platform.

• Windows — no setup; the PowerShell bridge spawns on demand.

• macOS — first run needs Accessibility (required) + Screen Recording (optional); clawdcursor grant walks the dialogs. Retina/HiDPI handled in-adapter — don't pre-scale coordinates.

• Linux X11 — apt install tesseract-ocr python3-gi gir1.2-atspi-2.0.

• Linux Wayland — same, plus ydotool + ydotoold (preferred) or wtype (keyboard only).

Safety & privacy

• Pauses instead of pushing through. At a sign-in, 2FA, CAPTCHA or plan-limit page the agent hands the desktop to you (blue glow, Resume / End task) instead of guessing. A pause times out after 15 minutes (CLAWD_PAUSE_TIMEOUT_MIN); CLAWD_NO_PAUSE=1 turns pausing off.

• Network isolation. Binds to 127.0.0.1 only; there is no setting that opens it to the network. Verify: netstat -an | findstr 3847 (Windows) / | grep 3847 (Unix). To drive it from another machine, tunnel the port over SSH: ssh -L 3847:127.0.0.1:3847 user@host, then use http://127.0.0.1:3847/mcp on your side.

• Bearer-token auth on every HTTP request (~/.clawdcursor/token).

• Sensitive-app policy. Email, banking, password managers, private messaging auto-elevate to Confirm.

• No telemetry by default. Nothing phones home. Screenshots stay in RAM; with a local model nothing leaves the machine; with a cloud provider, screenshots go only to the endpoint you configured. clawdcursor report is opt-in and previews exactly what it sends.

• No screenshots to any model, if you want. clawdcursor agent --no-vision (or CLAWD_DISABLE_VISION=1, which also covers stdio clawdcursor mcp) keeps every image from the built-in agent and from MCP clients; they work from the accessibility tree and OCR text.

• Prompt-injection defense. Screen text is returned inside <untrusted-screen-content> tags — data, never instructions.

• Log privacy. Logs redact password-field values (AXSecureTextField, UIA IsPassword=true).

See SECURITY.md for private vulnerability reporting.

Architecture

The PlatformAdapter is the only thing platform code talks to; safety.evaluate() is the only way tools execute. Those two seams are the whole point.

CLI

For humans diagnosing an install. Agents connect via MCP.

clawdcursor consent         Manage desktop-control consent (--accept / --revoke / --status)
clawdcursor grant           Grant macOS permissions (interactive, macOS only)
clawdcursor doctor          Configure the AI provider for `agent` mode (+ diagnostics)
clawdcursor status          Readiness check (consent, permissions, AI config)
clawdcursor versions        Every installed copy and its version (global, Claude Desktop extension, Claude Code plugin)
clawdcursor resume          Continue a task paused for you (same as the on-screen Resume); --end to stop it
clawdcursor mcp             stdio MCP server — editor hosts spawn this; you don't
clawdcursor agent           Daemon: HTTP MCP on :3847, optional built-in thin loop
clawdcursor agent --no-llm  Daemon, tool surface only (no built-in brain)
clawdcursor stop            Stop every running mode
clawdcursor uninstall       Remove all config and data

Options:  --port <n> (default 3847) · --compact · --no-banner · --provider <name> · --accept

Development

git clone https://github.com/AmrDab/clawdcursor.git && cd clawdcursor
npm install
npm run build       # tsc + postbuild  →  dist/surface/cli.js
npm test            # vitest (1,000+ tests)
npm run lint        # eslint
npm link            # global `clawdcursor` shim (Admin shell on Windows)

Tests run on Node 20 & 22 against Ubuntu, macOS, and Windows in CI, plus a coverage ratchet, a perf tripwire, and an npm audit gate.

Tech: TypeScript · Node 20+ · nut-js · Playwright · sharp · Express · Model Context Protocol SDK · Zod · commander.

Contributing

PRs welcome — see CONTRIBUTING.md for the dev loop, branch conventions, and the test matrix every change clears. Bugs and features in issues; private security reports via SECURITY.md.

License

MIT — see LICENSE.

Acknowledgments

Built on the Model Context Protocol SDK, nut-js, Playwright, the Anthropic computer_20250124 tool shape, and the AT-SPI / UIA / AX trees that make app-agnostic GUI automation possible at all.