Ghost

Verified control of the whole desktop, for agents and programs, on Windows and Linux. Ghost gives Claude Code, Codex, Cursor, any MCP harness, or a plain script the operating system's own control surface: the apps with no API, the windows, the shell, and the browser you are already logged into - in the background without taking your screen or cursor, with every action proven to have happened.

Like Playwright, but for the native desktop, and built for agents. Ghost does the perceiving, the acting, and the verifying; the model you are already running does any looking that is needed, so there is no vision API key to set.

One MCP surface, two engines: Win32 UI Automation on Windows, AT-SPI2 over D-Bus on Linux. The verbs, the locator tiers and the act-then-verify loop are written once and behave the same on both. Platform support · Linux setup

Why Ghost is different

• Runs in the background, and that is enforced. An agent can click, type, and use shortcuts inside an app while you keep working in another window - no focus steal, no cursor jump. It posts window messages to real controls and uses UI Automation patterns on windowless ones; most tools can only drive whatever is in the foreground. Since 0.19 this is a policy, not a preference: the focus policy defaults to background. Since 0.20 it is also constructive: anything Ghost starts (an app, a windowed browser) is born on a hidden desktop that has its own input queue and cannot take your foreground, and the ordinary verbs drive it there by window title. A call that truly has no background path fails naming the action instead of quietly taking the screen. Since 0.22 the policy is locked there: no tool call can raise it, so an agent cannot decide on its own to take your mouse. You can, by setting GHOST_FOCUS_LOCK=off in the server's environment. And since 0.23 a window that grabs the foreground on its own - browsers do, on their own accessibility calls - is handed straight back, so your typing keeps going where you are looking. Measured on a real desktop with a person typing throughout: of roughly 490 keystrokes a run, zero to one reached the window the agent was driving. (how)
• Never your window by accident. The session remembers the last window the agent named or launched and every window-scoped verb targets it by default. The human's foreground window is used only when nothing was ever anchored, and the response says so. Three weeks of real transcripts showed "element not found in the foreground window" as the top failure before this; it was the agent searching the window you had open.
• Many agents at once. Requests are dispatched in parallel, so a 15-second wait in one tab does not stall an instant query behind it, and a second Ghost process runs its own browser alongside the first without contending for the mouse.
• Prove it on your machine. ghost verify drives the real MCP server over stdio and audits every claim above against hard timing budgets, exiting non-zero if any of them does not hold on your hardware.
• Every action is verified. Ghost re-checks the screen (or reads the control's value back) after acting and returns verified / focus_confirmed - never a blind ok:true. Agents fail by acting and not knowing if it worked; Ghost closes that loop.
• Drives apps with no API. Legacy Win32, WPF, Electron, UWP, vendor portals - the software that has no integration and most needs automating. No CDP, no browser, no app cooperation required.
• No vision key. The model driving Ghost reads ghost_see (every element with its name, role and on-screen centre) or ghost_screenshot itself. Ghost's own vision tier is optional, exists for callers with no model of their own, and is off until configured.
• Accessibility-native and deep. Real element discovery through the OS's own accessibility API - UI Automation on Windows, AT-SPI2 on Linux - not pixel-guessing. Elements come back with real names, roles and bounds.

See it in one script: examples/background_agent_demo.py drives an app in the background while the foreground stays yours. Honest comparison vs Playwright-MCP / cua-driver / Computer Use: docs/comparison.md.

What people use it for

Vision is the smallest part of it. In a typical session an agent calls Ghost mostly to act, to manage windows and processes, to run commands, and to read state back.

Every row runs under the same guarantees: the action returns verified, the default policy never takes your foreground, and Ctrl+Alt+G stops every Ghost process at once.

What is Ghost?

Ghost is the layer between a model and the desktop. The model reasons; Ghost sees the screen the way the operating system does, acts on real controls without touching your foreground, and reports whether the action took. It gives programmatic control over any desktop application - native Win32, Electron, WPF, UWP, GTK, Qt, or otherwise - to an agent, a script, or a program.

On Windows it uses UI Automation for element discovery, SendInput for keyboard/mouse injection, and DXGI/GDI for screen capture. On Linux it uses AT-SPI2 over D-Bus for discovery and actions, XTEST (X11) or the RemoteDesktop portal / uinput (Wayland) for input, and X11 GetImage or the Screenshot portal for capture. The Linux engine is pure Rust - no -devel packages to install.

Ship it three ways:

• ghost-mcp server - the primary surface: a Model Context Protocol server for Claude Code, Claude Desktop, Codex, Cursor, and any MCP client (54 tools on Windows)
• ghost CLI - one-shot commands, great for scripts and CI (ghost click --name "Submit")
• ghost-http server - local REST API, call it from Python, Node, curl, anything (curl http://127.0.0.1:7878/list-windows)

The MCP surface is 20 desktop verbs, 19 ghost_browser_* / ghost_tab_* tools for driving individual browser tabs in the background (Chrome, Comet, Edge, Brave), and 15 Windows-only tools: the focus policy plus ghost_desktop_* for explicit control of isolated Windows desktops the user never sees. Under the default policy you rarely need the latter: ghost_window op=launch already starts the app on the hidden desktop auto, and ghost_see / ghost_act / ghost_key / ghost_scroll reach it with window=<title> exactly as they reach a window on your own desktop (target.surface in the response tells you which). UIA, window messages and capture work fully there. Real SendInput does not, because Windows refuses it off the input desktop, and typing is proven by reading the control's value back, so a target that drops posted characters returns an error rather than a false success. The desktop verbs and the browser tools build on Linux as well; the focus policy and hidden desktops are Windows-only.

No Claude required. No browser required. No CDP. It drives apps through the OS's own automation and input APIs, so it works with native apps that have no API and no automation hooks of their own - the same reliability whether or not an app was built to be automated.

Platforms

Wayland deserves the separate row rather than a footnote: it is the default session on current Ubuntu and Fedora, so it is what most Linux users would actually run, and it is the part with no test behind it. The discovery and action layer is shared with X11 and is covered, and it is the layer that matters most here - AT-SPI2 asks the application to do the thing, so there is no pointer to move and no window to raise. What is unverified is the fallback underneath: portal input, portal capture, uinput. If you run Wayland, treat ghost doctor as the first thing to run and expect to file bugs. Reports are welcome and are the fastest way that row changes.

ghost-session and ghost-mcp are shared: the locator tiers, grounding cascade, act-then-verify loop and the 20 core MCP verbs are written once and run on both platforms. Only the engine underneath changes, behind a one-line cfg alias. The browser and tab tools are engine-independent and build for both; the focus policy and isolated desktops are Windows-only and are reported as such rather than faked.

The wedge survives the port. On Windows, driving an app without stealing focus is built on posted window messages. Linux has a cleaner analogue in AT-SPI2 actions: the application performs the operation through its own toolkit, so there is no pointer to move and no window to raise. That layer is the same code under X11 and Wayland; synthetic input is only the fallback beneath it, and the fallback is the part Wayland has not been tested on.

This is tested, not asserted: CI stands up a real desktop (Xvfb + D-Bus + at-spi-bus-launcher), drives a real GTK application, and requires that text written through AT-SPI reads back from the app and that invoking a button actually dismisses the dialog. Wayland portal input and capture are implemented but not yet verified on hardware.

Linux setup, verification checklist and honest limitations: docs/linux-fedora.md. Capability matrix across all three: docs/cross-platform.md.

Ghost is a general-purpose automation tool. Use it on systems you own or are authorized to automate, and in line with the terms of the software you drive.

Install

One-click - MCP Bundle (free). Every release ships ghost-windows-x64.mcpb and ghost-linux-x86_64.mcpb on the Releases page. Open one in a client that supports MCP Bundles (Claude Desktop: Settings -> Extensions -> Install from file) and Ghost is registered, no PATH or config editing, and no API key. Ghost is also listed in the MCP registry as io.github.NORTHTEKDevs/ghost, so registry-aware clients can install it from there. The bundle holds the ghost-mcp server only; the CLI and HTTP server are in the archives below.

Prebuilt binaries (free). Every release ships signed-by-checksum archives for both platforms on the Releases page:

# Linux x86_64
curl -LO https://github.com/NORTHTEKDevs/ghost/releases/latest/download/ghost-linux-x86_64.tar.gz
curl -LO https://github.com/NORTHTEKDevs/ghost/releases/latest/download/ghost-linux-x86_64.tar.gz.sha256
sha256sum -c ghost-linux-x86_64.tar.gz.sha256
tar -xzf ghost-linux-x86_64.tar.gz && ./install.sh

Windows: download ghost-windows-x64.zip from the same page. Verify the checksum, unzip, and add the folder to your PATH. Then run ghost doctor.

Check where a download came from. Every release artifact carries a signed build provenance attestation, so you can prove a file came out of this repository's release workflow and nowhere else, at a named commit:

gh attestation verify ghost-windows-x64.mcpb --repo NORTHTEKDevs/ghost

The binaries are not code-signed yet, so Windows SmartScreen will warn you on first run (click More info → Run anyway). That warning is about publisher identity, which needs a paid certificate tied to a verified legal entity; the attestation above is the stronger statement about origin and costs nothing, but Windows does not read it. The pipeline signs the moment a certificate is configured, from any CA - see docs/code-signing.md, and docs/signing-policy.md for what gets signed, by whom, and what leaves your machine (nothing, unless you configure a vision key). If an antivirus engine quarantines a release, verify the checksum and see docs/antivirus.md for what the binaries do to stay recognisable and how to report a false positive. The kit buys convenience, not capability - everything Ghost can do is in the free source below, and building it yourself takes one command.

Option C - Build from source (free, MIT). Ghost is open source. Compile it yourself:

git clone https://github.com/NORTHTEKDevs/ghost
cd ghost
cargo build --release --bin ghost --bin ghost-http --bin ghost-mcp
# binaries in target/release/

Requirements: Windows 10 build 19041+, or Linux with at-spi2-core (and Rust stable only if building from source).

On Linux:

sudo dnf install at-spi2-core xdg-desktop-portal xdg-desktop-portal-gnome
gsettings set org.gnome.desktop.interface toolkit-accessibility true
./scripts/install.sh          # build, install, register the MCP server, run doctor

No -devel packages are needed - the Linux engine is pure Rust. Full setup and troubleshooting: docs/linux-fedora.md.

Check your machine first:

ghost doctor

Reports PASS/WARN/FAIL and exits 1 if anything is FAIL. Run it before opening an issue - it usually names the problem outright.

• Windows: build version, interactive desktop, UI Automation, DPI awareness, monitor layout, screen capture, optional vision credentials.
• Linux: session type (X11/Wayland), AT-SPI bus reachability, whether applications are actually exposing accessible trees, the selected input backend, and screen capture.

Quick Start - coding agents and MCP clients

This is the path Ghost is built for. Nothing to configure and no API key.

Claude Desktop: download ghost-windows-x64.mcpb (or the Linux bundle) from the latest release and open it: Settings -> Extensions -> Install from file.

Claude Code:

claude mcp add ghost --scope user -- C:/path/to/ghost-mcp.exe

Any other MCP client (Codex, Cursor, a custom harness): register the binary as a stdio server.

{
  "mcpServers": {
    "ghost": { "command": "C:/path/to/ghost-mcp.exe" }
  }
}

Ghost is also listed in the MCP registry as io.github.NORTHTEKDevs/ghost for clients that install from there.

The repo's Dockerfile builds a headless image (docker build -t ghost-mcp .) that answers initialize and tools/list with no display; registries and CI use it to introspect the server. A container has no windows to drive, so it is not an install path for real use.

How an agent uses it. The loop is look, act, confirm:

1. ghost_see window="Invoice Editor" - every element in the window with its name, role, enabled state, and on-screen centre. Text mode extracts the readable text instead, roughly ten times cheaper in tokens than an image.
2. ghost_act window="Invoice Editor" name="Save" action="click" - Ghost drives the control in the background and returns verified: true only when the screen or the control's value shows the action took.
3. ghost_screenshot window="Invoice Editor" - pixels, for the moments a layout, a chart, or a canvas needs the model's own eyes.

The model reads step 1 and step 3 and chooses; Ghost never has to guess what a picture means, and no vision key is involved.

Beyond the loop: ghost_shell runs commands and persistent PowerShell sessions, ghost_window launches, lists, focuses, restores and closes windows (on a hidden desktop by default), ghost_assert and ghost_wait turn "did it work" into a check, and the ghost_tab_* tools drive the browser you are already signed into. Most real sessions spend more calls there than on looking.

54 tools on Windows (legacy names stay dispatchable): 20 desktop verbs covering see/snapshot/find/act/keys/scroll/drag/clipboard/screenshot/windows/shell/waits/query/run, 19 ghost_browser_* / ghost_tab_* tools, and 15 Windows-only tools for the focus policy and isolated desktops. Building from source instead of downloading: cargo build -p ghost-mcp --release.

Every tool runs on its own task, so a slow call does not block a fast one, and a second Ghost process can run alongside the first. Once it is mounted, run ghost verify to audit that on your own machine.

Speed. Ghost's own time is small: about 75 ms for a text read, 200 ms for a verified click, 1 ms to list windows. What makes an agent session slow is what happens around the calls, and three habits remove most of it. Name the window once and omit window= afterwards: the anchor follows that window by handle, and a title it used to have still resolves instantly with title_drift in the response. Wait for a condition, not a duration: ghost_wait for=element, for=value, and for=navigate (sets the address bar in the background and returns when the title changes, about 0.4 to 1.4 s where a fixed sleep costs 6) return the moment the thing happened. Batch with ghost_run so one model turn does several steps. scripts/speed-probe.mjs measures all of this against any ghost-mcp binary.

Shell control (ghost_shell)

Ghost drives GUIs and the command line. ghost_shell runs terminal commands and persistent PowerShell sessions - builds, git, CLIs, file edits on hosts without file tools, or launching apps. op=run is a one-shot (powershell/pwsh/cmd); op=open starts a persistent PowerShell whose variables and cwd survive across op=send calls. Output is merged stdout+stderr, tail-capped for the agent's context window; a timed-out command keeps running and is drained with op=read; ghost_stop kills a runaway. op=run with the default powershell is served from a pre-spawned spare process, so a command costs about 85 ms instead of the 230-450 ms a fresh PowerShell start takes (the spare is single-use and replaced immediately; GHOST_SHELL_WARM=off disables it).

Spawn a fresh Claude Code session from the agent: ghost_shell op=run cmd='Start-Process wt -ArgumentList "pwsh","-NoExit","-Command","claude"', then drive the new terminal window with ghost_see / ghost_act / ghost_key.

Security: shell access is powerful. Set GHOST_SHELL=off in the server's env to disable the verb entirely - every op then returns a clear refusal, leaving the GUI automation verbs fully usable.

Quick Start - CLI

# Launch Notepad and type into it
ghost launch notepad.exe
ghost focus-window "Notepad"
ghost type --role edit --text "hello from ghost"

# Keys and hotkeys
ghost press Enter
ghost hotkey --mods Ctrl --key s

# Screenshot
ghost screenshot --out shot.png

# Enumerate windows or UI
ghost list-windows
ghost describe --window "Notepad"

# Click at coords or by name
ghost click-at 500 300
ghost click --name "Save"

# Run a JSON intent (finite-state machine with retries, timeouts, conditions)
ghost run my-flow.json
echo '{"ops":[{"op":"launch","exe":"notepad.exe"}]}' | ghost run -

Everything outputs JSON for easy piping into jq or scripts.

Quick Start - HTTP Server

Start the server:

ghost-http --addr 127.0.0.1:7878

Then from any language:

# Bash / curl
curl http://127.0.0.1:7878/list-windows
curl -X POST http://127.0.0.1:7878/click \
  -H 'content-type: application/json' \
  -d '{"name":"Submit"}'
curl http://127.0.0.1:7878/screenshot -o shot.png

# Python
import requests
requests.post("http://127.0.0.1:7878/launch", json={"exe": "notepad.exe"})
requests.post("http://127.0.0.1:7878/type",
              json={"role": "edit", "text": "hello from python"})

// Node
await fetch("http://127.0.0.1:7878/hotkey", {
  method: "POST",
  headers: { "content-type": "application/json" },
  body: JSON.stringify({ mods: ["Ctrl"], key: "s" }),
});

Endpoints: /health, /tools, /click, /click-at, /type, /press, /hotkey, /screenshot, /launch, /list-windows, /focus-window, /window-state, /describe, /clipboard (GET/POST), /run.

Quick Start - Rust SDK

[dependencies]
ghost-session = { git = "https://github.com/NORTHTEKDevs/ghost" }

use ghost_session::{GhostSession, By, session::Region};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let session = GhostSession::new()?;
    session.launch("notepad.exe").await?;
    let edit = session.find(By::role("edit")).await?;
    edit.type_text("hello world")?;
    let png = session.screenshot(Region::full()).await?;
    std::fs::write("screen.png", png)?;
    Ok(())
}

Reliability Model

Desktop automation driven from an MCP client has a hostile focus environment: between tool calls, the client's own terminal usually retakes OS focus. Ghost is built for that:

• ghost_act is atomic - find → act → verify via screen delta. One call, no cross-call race. Under the default background policy it drives the control in place and never raises the window; raise the policy and it additionally brings the target's window to the foreground first (AttachThreadInput, confirmed).
• Every action response is honest: verified (did the screen actually change), focus_confirmed (was the right window foreground), and a warning when either is off - never a blind ok:true. Check verified before re-issuing an action.
• Nothing is left running. Browsers Ghost launches are in a kill-on-close job object: when the server ends, however it ends, they end with it (dies_with_server on the launch response). At startup the server also sweeps browsers abandoned by earlier servers and reports them in ghost_stats.orphan_sweep.
• Anchor to a window - ghost_see, ghost_find, ghost_act, ghost_key, ghost_click_at, ghost_scroll, ghost_wait, ghost_assert and ghost_screenshot (unless full=true) all take window (a title substring, resolved across your desktop and Ghost's hidden desktops: exact title beats prefix beats substring, a window that is not minimised wins ties). The match becomes the session anchor: later calls without window target it, never the window the human happens to be using. ghost_window op=focus anchors without raising under the background policy; op=anchor sets, clears or reports it; every response carries target {hwnd, title, surface, source}. A title that matches nothing lists the open windows and returns code -32007.
• Your own browser, through its own protocol - when a window's process was started with --remote-debugging-port (Comet, Chrome, Edge, Brave), the same anchored verbs route through the DevTools protocol instead of UI Automation: DOM names and aria-labels rather than a sparse tree, selectors that survive re-renders, trusted input events into that renderer, full modifier combos, and focus emulation so pages that check document.hasFocus() still accept typing. Nothing about it can reach your foreground. The response carries route: {browser, port, tab} and coords: viewport; a browser without a port keeps the UI Automation path unchanged. GHOST_CDP_ROUTE=off turns it off. Start Comet or Chrome once with --remote-debugging-port=9333 to get it.
• Misses name the alternatives - "element not found" is followed by the closest element names in that window, so an agent does not spend a round trip on ghost_see to learn what the app calls the thing.
• Disambiguate duplicates - index selects the nth match when several elements share a name/role (multiple "Close Tab" buttons); responses carry a matches count.
• It audits itself - an independent sampler watches the foreground window and the OS's last-input time; any foreground change with no real hardware input behind it is recorded as synthetic, with the tool calls that were in flight. ghost_stats reports the tally, so the headline claim is proven continuously, not once.
• Read, don't screenshot - ghost_see mode=text extracts a window/page's readable text straight from the accessibility tree: faster and ~10x cheaper in tokens than images.
• Latency is visible: every response carries ms, and escalated: true flags when a find had to pay a network VLM round trip (local tiers: cache → UIA → OCR are all on-device).
• Windows never disappear: minimized windows stay in ghost_window list (with state) and op=focus auto-restores them. A window that something hid outright (not visible, not minimized) shows up with op=list include_hidden=true as state: "hidden", and op=state state=restore brings it back without activating it. op=state also reaches windows on Ghost's own hidden desktops, so an app started with op=launch can be closed by name.
• Stop always works: ghost_stop preempts the in-flight call the moment it arrives (dedicated stdin reader), and Ctrl+Alt+G remains the OS-level kill switch.

Background mode (agent-harness / computer-use)

Agent harnesses (OpenClaw, Hermes/cua-driver, and any MCP client) mount a computer-use tool to let an LLM operate the desktop. Ghost is that tool - and it acts without stealing your focus or moving your cursor, so an agent drives an app while you keep working in another window.

Since 0.19 this is the default and it is enforced. The process-wide focus policy starts at background, and every primitive that could only work by taking the real cursor or foreground window is gated behind it. There is no silent fallback: a call with no background path returns an error naming the action and the route that needs no policy change (launch the app on the hidden desktop, or the browser over CDP).

Since 0.22 the policy is also locked there. A default is only a promise if the agent cannot flip it, and until 0.22 any agent could call ghost_set_focus_policy foreground the moment a background action was refused, which is exactly when a human notices Ghost moving the mouse. Now that call is refused too, with an error that says so. The key is the operator's, not the agent's: set GHOST_FOCUS_LOCK=off in the MCP server's environment (the host config you write) to allow prefer_background and foreground again, and optionally GHOST_FOCUS_POLICY to start there. ghost_focus_policy reports the policy and whether it is locked; the server logs both at start.

// Only if you WANT an agent to be able to drive your real mouse and keyboard:
{ "mcpServers": { "ghost": {
    "command": "ghost-mcp",
    "env": { "GHOST_FOCUS_LOCK": "off", "GHOST_FOCUS_POLICY": "background" }
}}}

What you are agreeing to

Ghost has two settings that decide how much of your computer an agent gets. Installing the bundle shows both as checkboxes; running the binary directly, they are environment variables. The server prints where it stands on both at start-up, so the host's log always answers the question.

The shell is on by default deliberately. It is the capability, not a bonus, and a tool that quietly ships without the thing people install it for is worse than one that tells you plainly what it can do. If that is more than you want to hand over, the checkbox is right there and the refusal is explicit.

A locked policy still leaves one thing outside Ghost's control: an application can activate its OWN window, and Chromium does exactly that when an agent types into a page or clicks a button through the accessibility API. Nothing outside the browser can stop that call. So since 0.23 Ghost undoes it. The interference audit doubles as a sentinel: it is told the moment any window takes the foreground (a Windows event, about a millisecond) and hands it straight back unless you chose that window yourself - by clicking it or alt-tabbing to it, which it can tell apart from typing because it watches for REAL input, not synthesized input. ghost_session_state reports how often that happened (foreground_handed_back).

How well it works, measured with an independent observer while a person typed into another window for the whole run (scripts/background-desktop-probe.mjs, scripts/interference-watch.ps1):

Since 0.23.2 the foreground is watched by event rather than by polling, so the hand-back starts about a millisecond after the window takes it. Six clean runs with a person typing throughout gave 0, 0, 0, 1, 0 and 0 keystrokes into the window Ghost was driving. It is not a guarantee - undoing an activation can never be one - and for a guarantee let Ghost start the app, so the activation never happens at all.

For zero activation rather than a brief one, let Ghost start the app: anything it launches lives on a hidden desktop with its own input queue and cannot take your foreground at all, and a browser it launches is driven over CDP.

// Drive an app while the human keeps working. No flag needed: background is the default.
ghost_window { "op": "launch", "exe": "notepad.exe" }
// -> { "surface": "hidden", "desktop": "auto", "window": { "title": "Notepad", ... }, "target": {...} }
ghost_act { "role": "edit", "action": "type", "text_input": "hello" }   // targets the anchor
// -> { "verified": true, "focus_preserved": true, "cursor_preserved": true, "mode": "hidden" }
ghost_act { "window": "Comet", "name": "Post", "action": "click" }      // a window on YOUR desktop
// -> { "verified": true, "focus_preserved": true, "cursor_preserved": true, "mode": "background" }

• Launches never surface. Measured on this repo's CI box: Edge and Chrome activate their first window on launch in every launch style (normal, hidden, minimised, from a background parent, placed at -32000,-32000). A window created on your desktop takes your keyboard the moment it exists. So under the background policy Ghost never creates one there: ghost_window op=launch, ghost_run launch steps and ghost_browser_launch mode=windowed start on a hidden desktop with its own input queue, and the window is anchored so the next ghost_see shows it. An independent observer sampling the foreground at 100 ms across a full launch-drive-close run reported zero changes.
• True background via posted window messages. Real Win32 controls are driven with BM_CLICK / WM_LBUTTONDOWN·UP (click), WM_SETTEXT (type) and WM_MOUSEWHEEL (scroll). These do not activate the window.
• Windowless controls, without the screen. UWP/WinUI/Chromium/Electron controls have no window handle. Ghost drives them with UI Automation patterns - Invoke for a click, ValuePattern for typing. Chromium answers some of these by activating its own window when Windows lets it (right after you used that browser): measured on Edge, a SetValue on a web input or the address bar, an Invoke on a page button and a posted click all pulled the window forward within ~90 ms. Since 0.21.10 that is undone on the spot: Ghost watches the foreground around every user-desktop verb, and when the target or any window of its process takes it, hands it back to the window you had, in 30 to 50 ms, and reports focus_preserved: false with a focus_guard record saying so. Posted single keys reach the page's focused element and do not activate anything.
• Shell children never open a terminal. The MCP server has no console, so a shell started plainly has none either, and every console program it ran (node, python, cargo, cmd) got a brand-new Windows Terminal window that took the foreground for the length of the run. ghost_shell now creates its shells with an invisible console; children inherit it and no window ever appears. Measured: nine foreground changes in seven seconds before, zero after, output unchanged.
• Verified even while occluded. type is confirmed by reading the control's value back; click by a PrintWindow before/after delta that renders a window that isn't visible. Every response carries verified, focus_preserved, cursor_preserved - Ghost never claims a background action it can't confirm.
• Hidden desktops are the same vocabulary. A window on the hidden desktop is driven with the same window=<title> calls; target.surface: "hidden" is the only difference. Chromium and Electron windows there are driven by posted messages rather than UIA actions (Chromium services Invoke/SetValue on a non-composited desktop only after a ~2 s internal wait; a posted click lands in ~100 ms), and pixel verification is skipped for them because a software render there costs seconds - confirm through ghost_tab_eval or ghost_see. Real SendInput still does not work on a hidden desktop (Windows refuses it off the input desktop), so an app that answers only real hardware input needs the foreground policy on your own desktop.
• Honest about single-instance apps. Windows 11 Notepad, Explorer, or a browser on its default profile hand a launch to their already-running process, whose windows live on your desktop. Ghost cannot prevent that; it reports surface: "user" with a warning rather than claiming the app is hidden.

Supports click, type, double_click, right_click, hover and scrolling, plus ghost_key for single keys (Enter/Tab/F-keys/char via WM_KEYDOWN/WM_CHAR). The background: true flag is accepted for compatibility; it is the default behaviour.

The clipboard and edit combos work in the background too - Ctrl+C, Ctrl+X, Ctrl+V, Ctrl+Z and Ctrl+A are sent as the semantic messages an app actually implements (WM_COPY, WM_CUT, WM_PASTE, WM_UNDO, EM_SETSEL) rather than as a posted modifier that apps reading GetKeyState would ignore. Combos outside that set are rejected rather than silently dropped, because posting cannot set the modifier state those apps read; use the foreground policy for them.

Vision: your agent is the model

Ghost is eyes and hands, not a brain. When a model drives it over MCP, that model does the looking: ghost_see gives it the window as structured elements with coordinates, ghost_see mode=text gives it the readable text, and ghost_screenshot gives it pixels when the structure is not enough. Claude, GPT, Gemini, and the strong open-weight vision models all read those directly. There is no key to set and nothing to configure, and Ghost never has to interpret an image on the model's behalf.

The optional built-in vision tier exists for callers that have no model of their own - the CLI, the HTTP API, an intent file - and for the convenience of a description target (ghost_find description="the blue Submit button") when you would rather have Ghost resolve it than read the tree yourself. It works with any tool-capable vision model behind an OpenAI-compatible endpoint (OpenAI, Gemini, Groq, NVIDIA, or a local vLLM / Ollama / LM Studio server) or Anthropic. Point GHOST_VISION_BASE_URL and GHOST_VISION_MODEL at the endpoint and set GHOST_VISION_API_KEY (a keyless local server needs only the base URL). Without it, description targets return a clear error naming the alternative; every other tool is unaffected.

Emergency Stop

Press Ctrl+Alt+G at any time to immediately halt every acting call. Read-only queries (ghost_see, ghost_snapshot, ghost_window list) keep answering, so you can still inspect what happened while everything is stopped.

• All queued actions are cancelled
• Any held modifier keys (Shift, Ctrl, Alt) are released immediately
• No stuck keys, no stuck modifier states
• Session-wide since 0.19. The stop is a named kernel event (Local\ghost-emergency-stop-event), so one press halts every Ghost process in your logon session, not just the one that happened to register the hotkey. ghost_reset resumes service, again for all of them.

Element Locators

session.find(By::name("Save")).await?          // by accessible name (substring)
session.find(By::role("edit")).await?          // by UIA control type
session.find(By::role("button")).await?

From the CLI: ghost click --name "Save" or ghost click --role button.

Intents - Declarative Flows

Write reproducible multi-step flows as JSON. The FSM executor supports retries, timeouts, and JSONLogic conditions for abort_if / retry_if.

{
  "ops": [
    { "op": "launch", "exe": "notepad.exe" },
    { "op": "focus_window", "name": "Notepad" },
    { "op": "type", "role": "edit", "text": "hello" },
    { "op": "hotkey", "mods": ["Ctrl"], "key": "s" }
  ]
}

Run with ghost run flow.json, POST /run, or ghost_execute_intent over MCP.

Architecture

ghost-cli     ghost-http     ghost-mcp     Rust SDK
    \            |            /   |           |
     \           |           /    |           |
      +-----> ghost-session <-----|-----------+   ← safe Rust API
              /           \       |
      ghost-core      ghost-linux |               ← one cfg alias picks the engine
          |                |      |
    Win32 UIA,       AT-SPI2 over |               ← ghost-core: SendInput, DXGI/GDI
    posted msgs      D-Bus, XTEST |               ← ghost-linux: portal / uinput
          |                |      |
      Windows OS        Linux     +-> ghost-browser  ← CDP over the DevTools port,
                                                        engine-independent

Supporting crates: ghost-cache (UIA snapshot + delta), ghost-intent (FSM + JSONLogic executor), ghost-ground (the locator tier cascade), ghost-platform (the capability matrix reported per OS).

The built-in vision tier (Set-of-Marks)

This section describes the optional tier above; an agent driving Ghost over MCP gets the same information from ghost_see and does not need it.

When you locate an element by natural-language description (ghost_find description="the blue submit button", or when a name/text lookup misses and escalates to the VLM), Ghost does not ask the model to guess pixel coordinates - models are unreliable at that (in testing, a plain "give me the coordinates of the equals button" landed ~250px off the target). Instead it uses Set-of-Marks: it overlays numbered badges on the window's detected elements, sends that marked screenshot plus each badge's accessible-name label, and asks the model which number matches. The number maps back to that element's exact rect, so the result is a real on-element coordinate, not a regression guess.

In a live check on Calculator, four descriptions ("the equals button", "the plus button", "the number seven key", "the multiply button") each landed exactly on the correct button - versus ~250px off with coordinate regression.

Honest scope: when detected elements carry accessible names (most apps), the labels do much of the disambiguation; for unlabeled icons the model leans on the badge's visual position/appearance.

Canvas / no-accessibility-tree apps. When the UIA tree is sparse (custom-drawn UIs, remote-desktop surfaces, game canvases), Ghost augments the Set-of-Marks candidates with a built-in CPU classical-CV detector (ghost_ground::cv_detect): edge density → connected components → size/aspect filter, no GPU and no model download. It gives the VLM real boxes to pick from where the accessibility tree has nothing. It is coarser than a trained detector - an optional OmniParser ONNX tier (--features yolo + GHOST_YOLO_MODEL) plugs into the same Set-of-Marks path when a GPU model is available. The CV-marks → VLM-pick end-to-end needs a configured vision key.

Benchmark - task success, not "did the call return ok"

bench/ holds a reproducible end-to-end benchmark: it drives the real ghost-mcp binary through 14 Windows desktop tasks and scores each by re-observing the actual result (does the Calculator display really read 42? is the typed value really present?), never by trusting a tool call's return.

Latest run (see bench/results/latest.md):

14/14 tasks passed (100%), median ~2.7 s per task (full wall-clock incl. app launch) - perception, click/keyboard action+verify, waits, window management (list/minimize/restore), text extraction, disambiguation, flow chaining, clipboard round-trip, structured errors, element screenshots, and value assertions.

And it proves it can fail: --self-test runs deliberately-wrong negative controls (assert the display reads 99 when it reads 42, etc.) and passes only if the harness scores every one as FAIL - so the green run above is a real signal, not a rubber stamp.

Reproduce on any Windows 10/11 machine:

cargo build --release -p ghost-mcp
python bench/run_bench.py             # exit 0 iff every task passed
python bench/run_bench.py --self-test # exit 0 iff the harness caught every planted failure

Reliability soak

bench/soak.py drives many act-then-verify cycles and gates on the signals unit tests can't see: how often verified comes back null/false, focus-loss rate, error rate, whether each action's real effect happened (the display is re-observed, never trusted from the return), and latency percentiles.

Latest (160 acts): PASS - verify-null 0.0, focus-loss 0.0, effect-mismatch 0 (100% correct), p50 85ms / p95 117ms. See bench/results/soak.md.

python bench/soak.py                  # exit 0 iff every reliability threshold holds
python bench/soak.py --cycles 250     # ~1000 acts
python bench/soak.py --self-test      # exit 0 iff the harness flags a planted-wrong effect

We deliberately publish only Ghost's own measured numbers - never invented columns for other tools. bench/README.md gives an honest protocol for comparing against Playwright-MCP / Computer Use / UI-TARS, and explains why a naive same-suite comparison isn't apples-to-apples (Playwright is browser-only; vision agents need an API + VM).

Microbenchmarks

End-to-end capture measurement (release, crates/ghost-core/tests/capture_latency_probe.rs) corrected the v0.10.0 assumption: the DXGI acquire dominates and hits a cliff on large windows, so region captures (act-verify, screenshots, Set-of-Marks) route through flat ~16.5ms GDI BitBlt in v0.11.0; full-screen still uses DXGI. Run the convert microbench: cargo bench -p ghost-core --bench convert. Older baselines: docs/benches/v030-baseline.md.

Requirements

• Windows 10 build 19041 or later
• Linux with at-spi2-core and a desktop session (X11 or Wayland); xdg-desktop-portal-gnome additionally for Wayland input and capture
• A Chromium-family browser (Chrome, Comet, Edge, or Brave) for the ghost_browser_* / ghost_tab_* tools only; the desktop verbs need none
• Rust stable (only for building from source)

License

MIT - Copyright 2026 Northtek