Your phone
Reads the Android screen, opens apps, taps, types, swipes.
One endpoint, nothing to install. Most assistants take a single line.
Add --scope user to make it available in every project instead of just this one.
Melaya asks which permissions to grant. Tick what you want, leave the rest dark, then just ask.
Create a free accountOn Claude Code or Codex the CLI is faster than any settings screen.
You get sent to a Melaya screen once, where you pick which permissions to grant. Then you are done.
Your assistant is smart and blindfolded. You are its eyes, its hands, and its clipboard.
You screenshot so it can see a screen it could read itself.
You copy and paste between the assistant and the app that does the work.
You do the clicking because it can only tell you where to click.
Screens were built for eyes and fingers. Your assistant has neither.
Three surfaces, one connection. Laya, Melaya's free System One model, finds loosely described targets for the fast phone and browser tools.
Reads the Android screen, opens apps, taps, types, swipes.
You attach the tab. It reads, acts, then verifies.
Hand off the jobs that outlast a conversation.
No syntax to learn. Just ask.
Build a Melaya pipeline agent that scans Product Hunt every morning, shortlists the launches that fit melaya.org, and drafts each maker a Gmail outreach tuned to my SEO. Let me approve the shortlist before it sends.
On my phone, go through my unread Instagram DMs, summarize who wants what, and flag the three worth answering today. Do not reply to anyone.
Open the supplier portal tab I just attached. Pull every invoice from August into a table with number, date and amount.
Every Monday at 8am, run the agent that checks last week's runs and tells me which ones failed and why.
Tick what you would grant. The tool list your assistant receives is cut to match, before it ever sees it.
Everything granted. Most people grant far less.
The short version: one address, permissions you pick, and an agent that can narrow its own access but never widen it.
A vendor-neutral standard that lets an AI assistant use outside tools, so one endpoint works in every assistant that speaks it.
Not for phone or account access. Paste the address and approve the connection. Browser control needs the Melaya extension in your own browser, and autonomous agents need the Melaya runner on your own machine.
Any MCP client. Claude Code, Claude, ChatGPT, Cursor, Codex, Le Chat, VS Code, Windsurf, Zed, Cline, Goose, Gemini CLI and Qwen Code all take the same endpoint.
No. Authentication is OAuth 2.1 with PKCE. You approve permissions on a Melaya screen and the assistant receives a scoped token. Provider and connector credentials are resolved server-side and never reach the assistant.
No. It can narrow its allow-list during a run and never widen it. Granting happens in the Melaya app or on your phone, on a screen that can ask you properly.
Phone control is Android only, version 8 or newer. There is no iOS build, because the accessibility access this needs does not exist there.
A Melaya account is free to create. The MCP endpoint is included. Autonomous agent runs use your own model subscription on your own machine.
Laya is an open Apache-2.0 System One model that answers typed questions with calibrated probabilities and zero generated tokens. Melaya runs it free on its own CPU: melaya_browser_fast and melaya_phone_fast use it to find loosely described targets, and pipelines you run use it through the decide tools. Read the Laya System One model guide on melaya.org for the details.
One address. One approval screen. Grant the scopes you want and leave the rest dark. Then stop being the mouse.
OAuth 2.1 with PKCE. Your password is never shared. The tool list is filtered to the scopes you approved.