Install it, connect once, then hand it work.
One click in the Microsoft Store. Free, signed by Microsoft.
Pick a level for each app. Try it here.
Plain words in, checked steps out. The important ones wait for you.
Copy this week's invoices from Outlook into Invoices.xlsx and draft a summary for Sam.
Same job, same 20 cells, same clock. One reads pixels and asks before every click. Melaya reads the window once, plans everything, then checks each step itself.
14.3× faster in this illustration: a 20-cell form at about 4 seconds per AI call. Melaya asks the AI again only when a step does not do what the plan said.
Other agents ask a big AI before every single click. Melaya asks once, then clicks, checks and keeps going on its own.
The agent planned every stroke as code, saved it as a plan, then started it with one call. Melaya moved the brush, switched colours and checked its own work. This is the real canvas.
melaya_computer_fastmelaya_computer_batchmelaya_computer_planThe same moves work in Excel, Outlook, Teams, File Explorer, Finder, SAP GUI, a CRM or an in-house tool. If you can do it with a mouse and a keyboard, an agent can do it under your rules.
screen · ocrUI Automation on Windows, Accessibility on macOS, AT-SPI2 on Linux: named elements with stable references. OCR and marked screenshots fill in for apps that draw their own UI.
click · type · pressMenus, combo boxes, checkboxes, drag and drop, Ctrl+S or Cmd+Shift+4. Values can be set directly instead of typed.
file_dialogFills file dialogs from file handles, never raw paths. A file it saves comes back to the run as a handle you control.
read_table · set_valueReads grids as rows in Excel, ERPs and CRMs, so an agent can loop over every open invoice instead of squinting at pixels.
delivery: backgroundWhere the OS allows, actions go through the accessibility layer: no cursor grab, no window jumping to the front. When it needs the foreground, it says so.
expect · observeAfter each write, Melaya looks at what changed. Nothing changed? The step reports no effect instead of a false success.
handoffSign-in, 2FA or a CAPTCHA? The agent pauses, you do that step yourself, press Continue, and it picks up from a fresh read.
keep_awakeKeeps the computer awake during a scheduled run, and says plainly when the screen is locked instead of clicking at nothing.
playbooksBuilt-in notes for 24 common desktop apps, from Excel, Outlook and Teams to Finder, with the keyboard routes that work best.
connectorsWhen Slack, Notion or Teams is connected, the agent uses the connector instead of clicking through the app.
Four ways to let an AI agent control a computer. All four are serious tools. Here is how they differ, side by side.
| What matters | Melaya | OpenClaw | Hermes Agent | Claude Cowork |
|---|---|---|---|---|
| How it sees the screen | Accessibility tree first (UI Automation, macOS Accessibility, AT-SPI2), OCR and marked screenshots as fallback, one reference space | Screenshots plus accessibility tree with element refs, Peekaboo on macOS by default | Set-of-marks screenshots, vision or accessibility-tree modes | Screenshots and coordinates, no accessibility tree given to the model, images downscaled to about 1372x887 |
| Model calls and token use | One model call plans many steps (up to 100), Melaya runs and checks them | Users report 80 to 150k input tokens per task | Not documented | Reviewers found it slow and token-hungry (a 5-hour Pro allowance used in 30 minutes, PCWorld) |
| Verified writes, no false success | Every write checked against the window, "no effect" instead of a false success, admin windows reported up front | Not documented | Documented: on elevated Windows windows, capture returns 0 elements and click reports success while doing nothing | Not documented |
| Background control on Windows | Background first through accessibility patterns, says when it had to use the foreground | Windows and Linux driver marked experimental, primary display only | Background-first by design | Works on the visible screen on Windows (background windows on macOS 15+) |
| Linux and headless VMs | Linux on X11 and Wayland, plus headless servers and VMs with npx @melaya/runner --computer | Windows and Linux driver marked experimental | Not documented | Not documented |
| Per-app permission levels | View, click or full control per app, sensitive apps off by default | No per-action confirmation, safety through durable arming and tool policy | Approval modes, including a bounded mode | Strong per-app approval tiers |
| Stop and safety | Stop button and shortcut, pauses when you touch the mouse or keyboard, never types into password fields | CVE-2026-25253 (one-click RCE) and malicious skills found in its skill hub | Relies on private macOS APIs for background input, which can break with OS updates | Esc stops it from anywhere, one session at a time (machine-wide lock) |
| Phone, browser and computer in one | Yes: the same agents drive your Android phone, your browser and your computer | Not documented | Not documented | Not documented |
| Scheduled runs on your computer | Scheduled pipelines run on your own computer, kept awake for the run | Cron and heartbeat scheduling | Not documented | The machine must stay awake for local apps |
| Any model, MCP access | Claude, GPT, Gemini, Qwen or local models, and reachable from Claude Code, Codex and Cursor over MCP | Not documented | Not documented | Runs on Claude, a very capable model |
Competitor details per public documentation, October 2026. "Not documented" means we found no statement in their public docs, not that the feature is missing.
OpenClaw is open source, with a huge skill ecosystem and cron and heartbeat scheduling. Per its public docs, its Windows and Linux driver is experimental and primary-display only, and users report 80 to 150k input tokens per task. Melaya runs on Windows, macOS, Linux and headless VMs, plans many steps per model call, and checks every write.
OpenClaw alternative: full comparisonHermes Agent from Nous Research is background-first, with a learning loop that writes skills and approval modes including a bounded one. Its own docs note that on elevated Windows windows, a click can report success while doing nothing, and that background input on macOS relies on private APIs. Melaya detects admin windows up front and reports no effect instead.
Hermes Agent alternative: full comparisonClaude Cowork brings a very capable model, strong per-app approval tiers and Esc to stop from anywhere. It works from screenshots and coordinates, uses the visible screen on Windows and runs one session at a time. Melaya reads the accessibility tree, works in the background where the OS allows, plans many steps per call, and can run Claude or any other model.
Claude Cowork alternative: full comparisonOpenClaw, Hermes Agent and Claude Cowork are trademarks of their respective owners. Melaya is not affiliated with them.
Agents only see and touch the apps you switch on, at the level you pick. These rules hold on every run, from every surface.
Checked twice: the server applies your levels, then the app on your computer checks again before it acts.
Linux is a first-class desktop here, not an experiment. The desktop app reads Linux apps through AT-SPI2 on X11 and Wayland. No screen at all? Run the Melaya runner on a headless server or VM and it becomes a computer your agents can use, with the same permissions, approvals and timeline.
Runner CLI and VM guidenpx @melaya/runner --computerOne set of computer tools, with the same app levels, approvals and Stop everywhere you use them.
Type the job in plain words and watch it happen, with approvals inline.
Scheduled runs on your own computer, from the tray, while you sleep or sit in meetings.
Claude Code, Codex, Cursor or Claude get computer tools with your permissions and approvals.
Works next to Browser Control and Device Control: read the web, update the desktop app, confirm on the phone, all in one run.
Five jobs teams hand to Computer Control first.
Clean a monthly export, fill formulas down, pivot by region and save a CSV where finance expects it. The agent reads cells as rows, not pixels, so it works through the whole sheet.
ExcelSort the inbox by sender and topic, file invoices, flag what needs you and draft replies for your approval. Sending always asks first in Safe mode.
OutlookPull the month-end ledger out of a desktop ERP, reshape it in Excel and save it on the shared drive. Schedule it and it runs every month on your own computer.
Desktop ERP · ExcelCopy leads from a spreadsheet into the desktop CRM field by field, and check that every record saved before moving on.
Excel · Desktop CRMClick through your own app after each build, compare what appeared with what should, and report the exact step that broke. Run it on a Linux VM in CI.
Your app · Windows, macOS, LinuxShort version: the agent works in the apps you allow, checks every step, and stops when you say stop.
An AI agent that operates a computer through its screen, mouse and keyboard, the way a person does, so it can work in apps that have no API. Melaya's computer use agent reads the app's accessibility tree first, acts on exact elements and checks each result.
Windows 10 and 11, macOS 13 or later, and Linux desktops on X11 and Wayland. Headless Linux machines and VMs work through the Melaya runner, started with npx @melaya/runner --computer.
Screenshot agents make a model call and take a screenshot for every click. System One reads the window once, plans many steps in one call, and Melaya grounds and checks each step on the accessibility tree. Far fewer model calls and images per task.
Yes. OpenClaw is open source with a large skill ecosystem. Per its public docs (October 2026), its Windows and Linux driver is experimental and primary-display only. Melaya runs on Windows, macOS, Linux and headless VMs, with per-app levels, approvals and verified writes.
Claude Cowork has a very capable model and strong per-app approval tiers. Per Anthropic's docs (October 2026), it works from screenshots and coordinates, without an accessibility tree. Melaya reads the accessibility tree, plans many steps per call, and can use Claude or any other model.
Hermes Agent is background-first and has a learning loop that writes skills. Its docs (October 2026) note that on elevated Windows windows, a click can report success while doing nothing. Melaya detects admin windows up front and reports no effect instead of a false success.
Password managers, banking and trading apps, terminals and system settings are off by default, and the agent never types into password fields. You decide which apps it may use and how far it may go: view, click or full.
Any model you use in Melaya: Claude, GPT, Gemini, Qwen, local models through Ollama or LM Studio, or your own Claude Code or Codex subscription. Fast mode keeps model calls low, so smaller models do well too.
The app and every action run on your computer. What the agent reads from the apps you allowed goes to the model you picked. Pick a local model and it stays on your machine.
Browser Control drives web pages, Device Control drives an Android phone, Computer Control drives desktop apps. Same agent, same permissions and approvals, and one run can use all three.
Full computer use comparisons:OpenClaw alternativeHermes Agent alternativeClaude Cowork alternative
Install the desktop app, connect this computer, switch on the apps you want help with, and hand over the first boring job.
Runs on your computer. Apps stay off until you allow them. Stop is one shortcut away.