Melaya

[Melaya: assistant, agent builder and paired phone on one runtime]

Melaya pairs a phone to your account and gives an agent the same view of it a person has: it reads the screen through Android's accessibility tree, then taps, types, swipes, and moves between apps. No per-app integration, no vendor API. If you can use the app, so can the agent.

This is a remote Model Context Protocol server. MCP is a vendor-neutral standard, so one endpoint works everywhere:

https://api.melaya.org/mcp

Nothing to install, no SDK required.

Connect

claude mcp add --transport http melaya https://api.melaya.org/mcp

codex mcp add melaya --url https://api.melaya.org/mcp

{ "mcpServers": { "melaya": { "url": "https://api.melaya.org/mcp" } } }

code --add-mcp '{"name":"melaya","type":"http","url":"https://api.melaya.org/mcp"}'

{ "mcpServers": { "melaya": { "url": "https://api.melaya.org/mcp" } } }

Authentication is OAuth 2.1 with PKCE. You choose which permissions to grant on a Melaya consent page; the assistant receives a scoped token. Your password is never shared.

Then just ask:

Connect my phone and go through my unread Instagram DMs.

Give your AI the Melaya skill

This server gives your assistant the tools. The Melaya skill gives it the method: which tool to call first, how to connect services, build and validate a pipeline on a real run, set triggers and approvals, and read results. Install both and your assistant gets real work right the first time instead of guessing.

Any assistant (ChatGPT, Gemini, Cursor, …): paste this line.

Install the Melaya skill from https://github.com/melaya-labs/melaya-mcp/tree/main/skills/melaya and use it whenever I ask you to work with Melaya.

Claude Code: the Melaya plugin from this repo already bundles it, next to melaya-setup. Installing the plugin is enough. To use the skill on its own, copy it into your skills folder, then restart. It loads on its own when you mention Melaya.

git clone --depth 1 https://github.com/melaya-labs/melaya-mcp /tmp/melaya-mcp && cp -r /tmp/melaya-mcp/skills/melaya ~/.claude/skills/

Claude.ai and Claude Desktop: download skills/melaya as a .zip and upload it under Settings → Capabilities → Skills. Anywhere else, add SKILL.md as a project instruction or rules file.

What it can do

Operate your phone

Read the screen, open apps, tap, type, scroll, swipe, screenshot. You watch it work, and one control stops everything.

[The agent working, with a live step trace and a stop control]

Melaya ships navigation playbooks for common apps, so the agent arrives knowing where things are instead of exploring blindly.

[App playbooks: navigation map, stable control ids, canonical step sequence]

Operate a browser

The same read-act-verify loop on a desktop site, through the Melaya extension on Chrome or Edge. You attach the tab; the model never picks one. Several conversations, pipelines and the extension panel can work in the same browser at once, each in its own tab (up to 8): attach returns an agent id that the conversation passes on every browser call, and a tab another agent is using is never touched.

[A live browser session, driven step by step]

It can also debug the page it is on: network activity, console output, and a performance diagnosis that ranks causes with the file and the number behind each, rather than handing over a raw panel. Credential values are redacted at capture, before anything reaches the model.

Build and run agents

Create a project for the work (one per client pilot, say), list the template library and instantiate a validated template, or author a pipeline from scratch, validate it before saving, schedule it, and watch it run. Hand a long or recurring job to an autonomous agent on your own machine, on your own model subscription, that carries on after the conversation ends.

[The Melaya Agent Builder]

Read, and act through, your connected services

Mail, documents, your ERP. Read-only by default, structurally: without the separate write permission the write path is blocked in two independent places, so a write stays blocked even if Melaya's own tool catalog is out of date.

Grant melaya:connectors.write as well and the assistant can act for you there too: send an email, create or update a record or a file. That includes tools that move money or trade (payments, refunds, purchases, transfers, ad-spend changes, orders): the consent screen says so, discovery labels them, and your assistant should confirm the exact action and amount with you first. Every write runs at once, with no second prompt, and is audit-logged. LinkedIn session tools are never reachable over MCP, at any permission.

Manage projects

Create a project, rename it or edit its description, and delete it once it is empty. Delete is a dry run until confirmed, only the project's creator can do it, and a project that still holds pipelines, run history or other members is refused. Plan limits apply exactly as in the app.

Find out what happened

Runs, transcripts, tool traces, failure diagnosis, cost, and what your agents have learned across runs.

What it cannot do

Enforced by the device or by the Melaya server, not by instructions, so no prompt and no agent instruction can move them.

It only touches what you allow-list

Apps on the phone, origins in the browser. The agent can hand access back, narrowing the list or clearing it, but only you widen it: phone apps on your phone or in the Melaya app, and browser sites on the Melaya Browser Control page or by asking your assistant to add one (it then calls melaya_browser_allow_sites with that site; you can also ask it for every site). The skills tell it never to widen access on its own or because of something it read.

[The allow-list, in the Melaya app]

That asymmetry is deliberate. The agent reads text off your screen, and text can be written by anyone: a message, a comment, a web page. A boundary it could widen in response to what it reads would not be a boundary.

Payments on your phone always ask you

Direct phone control from your assistant runs in safe mode: publishing and paying stage an approval card, and you see the exact text before anything goes out. Phone agents and phone pipelines your assistant starts or saves keep the payment card in every mode: autonomous is stored and run as payments_only from here, which lets an agent publish without a card you asked it to skip, but still stops before every purchase or payment. Only you can set a fully unattended phone mode, in the Melaya app. In the browser, the autonomy you choose in the Melaya extension decides (safe asks before purchases, publishing and other consequential actions; payments only asks before purchases; autonomous asks nothing). Approvals reach you even when the phone is locked.

[An approval card, showing the exact text before it publishes]   [An approval on a locked phone]

There is a STOP control

On the phone overlay and in the Melaya app. It halts everything immediately, across every agent and every connected assistant.

Approvals are listed here, never decided here

The gate exists to put a human between an agent and a consequential action, and the caller here is a model reading untrusted content. You approve in the Melaya app or on your phone.

Password fields are excluded from screen reads.

LinkedIn is not available on this surface: automating it through a logged-in session breaks LinkedIn's terms, so the server refuses the LinkedIn tools, the LinkedIn app on your phone and linkedin.com in your browser.

Also deliberately absent, and enforced by tests rather than by convention: administration (no honest consent sentence exists for it), and reading or setting connector credentials (you connect services in the app). Two event-trigger fields are the exception, used only if you choose to hand a secret to your assistant instead of pasting it in the app: a provider's webhook signing secret and a stream source's key. Both are stored encrypted and never returned.

Permissions

Eleven scopes. You grant them individually.

The tool list your assistant receives is filtered to what you granted, so connecting for phone control alone shows 25 tools rather than all 130. If a capability seems missing, you declined it; reconnect and approve it.

[!NOTE] If you also use the Melaya SDK, "connectors" means something different there. In the SDK it is project credential storage. Here, melaya:connectors is reading data from services you already connected, and melaya:connectors.write is acting through them. This surface cannot read, set or delete a connector credential (the trigger signing secret and stream-source key above are the only secrets it accepts, and it never returns them).

Requirements

• A Melaya account — free at melaya.org
• An Android phone (Android 8 or newer) for phone control. There is no iOS build.
• The Melaya extension on Chrome or Edge for browser control
• For autonomous agents: Node 18+, Python 3.11+, and a signed-in CLI on the machine hosting the runner. The runner starts from one command pinned to a reviewed release (npx -y @melaya/[email protected] --token=...); your assistant shows it to you and runs it only after you say yes, or hands it to you to run

Privacy

Screen and page content read during a run goes to Melaya and to whichever model you selected, for the duration of that run. The privacy policy covers collection, retention and deletion.

One credential does cross the boundary, and it is worth naming: if you set up the optional local runner, the command you are given contains a runner token. It is valid for 7 days, revocable in settings or with melaya_runner_revoke, and unavoidable because the runner starts from a command line. On a hosted assistant that command appears in your conversation history. The only other secrets that can appear are a webhook signing secret or stream-source key you choose to give your assistant for an event trigger (stored encrypted, never returned); provider and connector credentials are resolved server-side and never reach the assistant.

Disconnecting in Melaya settings immediately revokes the connection's ability to renew itself. A token it already holds keeps working until it expires, at most one hour.

What's in this repo

The server itself runs as part of the Melaya platform; this repo is its public manifest, packaging and documentation.

The rest of the platform

This server is one surface onto Melaya. Its tool domains map to the products behind them, so if a scope is useful the product page explains what it reaches:

The browser extension is what Browser Control talks to, and ships for Chrome, Edge, Brave, Opera and Firefox.

Pricing starts at a $0 Sandbox tier for evaluating the platform during open beta: see plans.

Links

• Full documentation
• Melaya SDKs — nine languages
• Product overview · All products
• Privacy Policy · Terms
• Support

License

Apache-2.0, matching the Melaya SDKs.

It covers what is in this repository — the registry manifest, the plugin packaging, the skills and this documentation. It is not a licence to the Melaya service itself: using the hosted server at api.melaya.org needs a Melaya account and is governed by the Terms.