Skip to main content
Glama

nostrings

CI npm License: MIT Platforms: macOS, Android

Computer use for coding agents on your Mac and your Android phone: one JavaScript call per step, accessibility diffs instead of screenshots, and your screen stays yours. No strings attached.

Someone browses Wikipedia in Safari while an agent works Calculator and types the result into TextEdit in the background

const calc = await ui.getApp("Calculator");
await calc.pressKey("Escape");
await calc.typeText("6*7=");
await calc.getAXState();
Calculator — window 0 "Calculator" — image 460x816 — changes since last read (~ changed, + added):
~					8 scroll area id=StandardInputView;value:42
~						9 text "Edit field" value="42"

On the next call, calc is still bound. An agent writes a few lines per step instead of making one tool call per click, and reads what changed instead of a fresh picture of the window.

  • A script per step, not a tool call per click. Bindings and helper functions persist across calls, so one call can click, type, loop, and read.

  • Diffs, not screenshots. The default read is the accessibility tree as text, and later reads show only changes. Screenshots and pixel clicks are there when the tree is empty (Citrix, canvases).

  • Background by default. Keys, clicks, drags, and scrolls reach the app while you keep working in another one. Your cursor stays put, and a drawn cursor shows what the agent is doing.

  • Your phone too. ui.device() drives an Android phone or emulator over ADB from the same session. On Android 14+ apps run on a hidden screen, so the phone's own screen is left alone.

In an author-run test of six setups on 12 macOS tasks (2 runs each), nostrings had the lowest total cost, $0.57 against $1.41 to $5.67, and the fewest round trips and the most passes.

What to know first: the Mac side is tested on macOS 27 on Apple silicon. You need the Xcode command line tools to build a small helper app, and you grant it Accessibility and Screen Recording. Background input uses two private macOS calls, with a foreground fallback. Android is experimental: tested on emulators (API 34 to 37) and one Galaxy S24, and it needs only adb. Only Claude Code has been tested end to end. nostrings was called glovebox until 0.5.0.

Quickstart

Requires Node 20.10+ and the Xcode command line tools (xcode-select --install).

npm install -g nostrings
nostrings signing --create   # optional: macOS keeps the grants across helper rebuilds
nostrings setup              # builds the helper, walks through permissions
nostrings doctor             # every line should say ok
nostrings smoke              # a real end-to-end test on Calculator
claude mcp add --scope user nostrings -- nostrings mcp

Android only? Skip the helper. With a phone (USB debugging on) or an emulator showing in adb devices:

npm install -g nostrings
nostrings doctor             # the Android lines should say ok
nostrings smoke --android    # installs a small test app and drives it
claude mcp add --scope user nostrings -- nostrings mcp

setup builds Nostrings Helper.app into ~/.nostrings/ and asks you to turn it on under System Settings > Privacy & Security > Accessibility and Screen & System Audio Recording. Input Monitoring is optional and only enables Escape-to-stop. The helper is its own app, so one grant covers every agent and terminal. signing --create makes a self-signed certificate in your login keychain that signs only the helper. Without it, macOS treats each rebuild as a new app; see the upgrade note.

Works with: Claude Code (tested end to end). Codex, Grok CLI, OpenCode, Gemini CLI, and Cursor have documented configs that have not been exercised. Any agent that can run shell commands can use nostrings run. See docs/agents.md.

Related MCP server: Charlotte

Why not Claude Code's built-in computer use?

If it fits your setup, it is the easier choice: nothing to install, supported by Anthropic, and it keeps your terminal out of screenshots. nostrings is for the cases where it does not fit. From Anthropic's docs, checked 2026-11-03:

Claude Code built-in

nostrings

Your other apps

Hidden while Claude works

Stay visible; input goes to the target app in the background

Default read

Downscaled screenshot

Accessibility text diff

Who can use it

Pro, Max, or Ultra through claude.ai; not Team, Enterprise, Bedrock, Google Cloud, or Foundry

Anyone; it never talks to a model or an account

Scripted runs

Not in -p mode

The benchmark drives it unattended from a script; claude -p itself not yet tested

Sessions

One at a time, enforced by a lock file

No global lock; named sessions run side by side

Remembering approvals

This session only

Allow for This Session or Always Allow

Browsers

View-only

Full control once you approve the browser, which is more power and more risk

Claude Code's built-in tool was not in the benchmark below, so this table compares documented behavior, not measured results.

Benchmark

Twelve macOS tasks, six setups, two runs each, 2026-11-03. Same model (Claude Sonnet 5.5, medium effort), prompt, and agent loop for every setup; only the tool changes. Task prompts name no tool, and a separate check that doesn't use the tool under test decides each result.

Setup

Passed

Round trips

Images returned

Cost at list price

nostrings

24/24

128

21

$0.57

Codex cua_repl, driven by Claude

20/24

166

30

$1.41

Anthropic toolset, run by nostrings

22/24

187

179

$1.65

agent-desktop

22/24

204

22

$2.32

open-computer-use

21/24

244

385

$2.76

Peekaboo

20/24

359

183

$5.67

Median round trips and cost per task for each setup

Where nostrings did not win. It was not cheapest on finder-rename ($0.061 against $0.041 for Codex cua_repl). It did not have the fewest round trips on 3 of the 12 tasks, or the fastest time on 3 of the 12. On remote-desktop it took 36 seconds against 22 and 23 for two others.

How much to trust this:

  • I wrote the tasks and ran the test, on one Mac, in one sitting. Several tasks (bulk-entry, remote-desktop, coexistence) were designed around things nostrings set out to do.

  • With two runs per task, pass counts mean little: 24/24 against 20/24 is not statistically significant. The cost gap is the sturdier result, because it holds task by task: nostrings cost less than each other setup on at least 11 of 12 tasks.

  • The nostrings tested is the code released as 0.3.0. Result files say 0.2.0, because the version had not been bumped yet.

  • "Anthropic toolset" is Anthropic's API computer-use toolset, executed through nostrings's foreground route. It failed coexistence because of that route, so that result doesn't describe a shipping product. "Codex cua_repl" measures the tool shape under Claude, not Codex with its own model.

  • Peekaboo's first pass was rerun with foreground input allowed, which helped it. Its two bulk-entry failures were my $0.75 per-run budget cap stopping the run. agent-desktop ran through its CLI, not its MCP mode.

Method, per-task numbers, and raw results: bench/ and bench/results/v2/summary.md.

How it compares

Checked against each project's README and docs on 2026-11-03; the last three rows on 2026-11-19. "—" means I could not confirm it from those.

Platforms

Actions per call

State between calls

Default read

Diffed reads

Background input

Needs

nostrings

macOS, Android (experimental)

many (one script)

persistent JS session

accessibility text

yes

yes

Xcode CLT to build the Mac helper; adb for Android

Claude Code built-in

macOS (CLI)

—

—

screenshot

—

no, hides other apps

Pro, Max, or Ultra

Codex computer use

macOS, Windows

many (one script)

persistent JS session

accessibility text

yes

yes on macOS

ChatGPT app and plugin

lcu

macOS (Apple silicon), Linux

as Codex; it runs Codex's runtime

as Codex

as Codex

as Codex

—

ChatGPT app installed

agent-desktop

macOS, Windows, Linux

one per command, plus batch

refs tied to a snapshot

accessibility tree JSON

—

yes

Homebrew or npm

Peekaboo

macOS 15+

one per command

element IDs from a snapshot

screenshot or JSON UI map

—

yes

Homebrew or npm

open-computer-use

macOS, Linux, Windows

one per MCP call; CLI has sequences

—

app state

—

—

npm

computer-use-mcp

macOS 14+

one per call

element IDs

screenshot and tree per action

—

yes

notarized binary

cua-driver

macOS, Windows, Linux

—

—

—

—

yes

see its docs

native-devtools-mcp

macOS, Windows, Android, Chrome/Electron

one per call

—

—

—

yes on macOS

see its docs

OpenComputerUse

macOS, Linux, Windows, Android, iOS

one per call, plus steps batches

session ids

—

—

yes

see its docs

mobile-mcp

Android, iOS (no desktop)

one per call

—

—

—

n/a

npm

Pick something else if you need Windows or Linux (agent-desktop, cua-driver, lcu), iOS (mobile-mcp, OpenComputerUse), a notarized binary with no build step (computer-use-mcp), or want OpenAI's actual runtime and already have the ChatGPT app (lcu).

Safety

The first time a session touches an app, you get a dialog: Allow for This Session, Always Allow, or Deny. Nothing is blocked outright. Terminals, password managers, System Settings, messaging apps, and a few others always ask, even under allow-all, unless you run nostrings allow <bundle id>; the list is in SECURITY.md. A script that had to bring an app forward stops if you switch away. Scripts running in the background keep going while you use other apps, and Escape stops any script if you grant Input Monitoring. Every action is logged to ~/.nostrings/actions.log.

An approved app is approved for everything in it: allowing a browser lets the agent use every site you are signed in to. nostrings does not screen page content for prompt injection. Treat it like giving someone remote control of that app. Details: SECURITY.md.

Known issue: while a background app's pop-up menu is open, macOS sends your keystrokes to that app, so a word you type can land in it (details).

How it works

agent ──MCP stdio──▶ nostrings mcp ──▶ kernel (persistent JS session)
agent ──shell──▶ nostrings run ──▶ daemon ──▶ kernel
                                                │
                       ┌── Unix socket ─────────┴──────── adb ──┐
                       ▼                                        ▼
       Nostrings Helper.app (Swift: AX,       nostrings server on the phone
       CGEvent, ScreenCaptureKit, dialogs)    (Java, shell user: accessibility,
                                              input, hidden virtual displays)

Background input on the Mac uses two private macOS (SkyLight) functions, looked up at runtime with a foreground fallback; nostrings doctor reports which route you have. On Android, a small server runs with shell privileges over ADB. It is a file pushed to the phone, not an installed app. Everything else is public API. No OpenAI, Anthropic, or other vendor binaries.

Status

Platform

Status

macOS

Works. Built for macOS 14+; tested only on macOS 27 (Apple silicon).

Android

Experimental. Hidden-screen sessions on Android 14+ (API 34); the phone's own screen below that. Tested on emulators API 34 to 37 and a Galaxy S24.

Windows

Planned (UI Automation). See roadmap.

Linux

Planned (AT-SPI; Wayland limited). See roadmap.

Run for real, not mocked: nostrings smoke and smoke --full over MCP (batched steps, persistent bindings, diffs, pixel clicks, and background click, drag, and scroll with another app frontmost), smoke --macos (51 checks on a test app), smoke --android (20 checks on a test app), and Claude Code attached through nostrings mcp. Full list: docs/testing.md.

The API in one screen

  • ui.getApp(name) launches or finds an app without stealing focus. ui.listApps().

  • ui.device() is an Android phone with the same getApp and app methods: const clock = await ui.device().getApp("Clock").

  • app.menu(["File", "Export…"]), app.ocr() for screens the tree cannot describe, and ui.screenshot(), ui.clipboard, ui.notifications for the rest of the Mac.

  • app.getAXState() returns the accessibility tree as text; later reads are diffs. {frames: true} adds element boxes, {disableDiffing: true} the full tree.

  • app.getScreenshot() is the only way to get an image.

  • app.click(target), doubleClick, rightClick, scroll, drag, setValue, focus, selectText, performSecondaryAction.

  • app.pressKey("ctrl+a") uses xdotool names (super is Cmd). app.typeText("text\n").

  • app.waitForChange() and app.settle() wait for a late update.

  • A target is an element index from the tree or [x, y] pixels in the window image, which is the screenshot's pixel space.

Full reference: docs/api.md, or help() inside a session.

Docs

License

MIT

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      B
      maintenance
      Enables an agent to observe and operate desktop apps and browsers through an accessibility-tree-first view, with clicking, typing, scrolling, dragging and waiting tools. Screenshot handling, coordinate mapping and fail-closed safety rules (risk confirmation, deny-listed apps, stale-observation refusal) are enforced inside the server.
      9
      Apache 2.0
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables AI agents to browse and interact with web pages through a persistent headless Chromium session, returning compact structured page summaries, targeted element queries, and full detail only on request. This keeps token usage far below full accessibility-tree snapshots while still supporting navigation, forms, screenshots, and session control.
      4 npm
      -
    • A
      license
      Not graded
      quality
      A
      maintenance
      Enables AI agents to browse the web through a persistent headless Chromium session, receiving compact structured page orientations, targeted element queries, and full detail only when explicitly requested. This typed decomposition of landmarks, headings, interactive elements, and forms keeps token cost far below raw accessibility-tree snapshots on content-heavy pages.
      4 npm
      MIT
    • A
      license
      Not graded
      quality
      A
      maintenance
      Enables AI agents to navigate, inspect, and interact with real websites via a persistent headless Chromium session, returning compact structured orientations, targeted element queries, and full detail only on request rather than dumping raw accessibility trees. This cuts token usage by roughly 10–140x on content-heavy pages while still supporting clicking, typing, form filling, screenshots, dialog handling, and multi-tab sessions.
      4 npm
      MIT