# MCP setup

Part of the Oya Browser docs. All of it: https://oyabrowser.com/docs.md

## MCP Setup

One endpoint for everything you have. It starts browsers itself, so there is no id to look up first:

```
https://oyabrowser.com/mcp/pool
```

Thirty-two tools: twenty-nine that act on a page, round-robined across your browsers, plus `start_browser`, `stop_browser` and `pool_status`. Tab tools, `read_elements`, `read_console`, `read_network` and the browser-specific tools marked in [More tools](https://oyabrowser.com/docs#more-tools) are not on this endpoint; to use those, point at one browser instead, `/mcp/{BROWSER_ID}`, with the id from the [dashboard](https://oyabrowser.com/dashboard).

`/mcp/pool` answers `initialize` and `tools/list` without a key, so an agent can list the tools before you sign in. Calls return 401 until you do. For everything an agent needs in one page, point it at [oyabrowser.com/llms.txt](https://oyabrowser.com/llms.txt).

### Desktop app

Install the [desktop app](https://oyabrowser.com/#download), then click **Connect** next to your agent's name in the Oya window footer. Claude Code, Codex, Cursor and VS Code appear there when installed. Connect writes an MCP server named `oya-browser` at `http://127.0.0.1:47321/mcp` into that agent's config; Claude Code also gets Oya's tools allowed without prompts. Start a new session and the tools are there. No account, no key.

### Cursor

Add to `.cursor/mcp.json` in your project, or let the desktop app write it for you with Connect:

```
{
  "mcpServers": {
    "oya-browser": {
      "url": "https://oyabrowser.com/mcp/pool",
      "transport": "streamable-http",
      "headers": {
        "Authorization": "Bearer YOUR_API_KEY"
      }
    }
  }
}
```

### Claude Code

On a machine with no Oya app, add the cloud endpoint with Claude Code's own CLI:

```
claude mcp add --transport http oya https://oyabrowser.com/mcp/pool \
  --header "Authorization: Bearer $OYA_API_KEY"
```

Or install the plugin, which brings the MCP server and a skill that teaches the workflow:

```
claude plugin marketplace add OyadotAI/oya-browser
claude plugin install oya-browser@oya

# any agent that reads skills, without the plugin:
npx skills add OyadotAI/oya-browser
```

### Codex

Codex reads `~/.codex/config.toml`. This is the section the desktop app writes; swap the URL and token for the cloud endpoint and your key when there is no desktop:

```
[mcp_servers.oya-browser]
url = "https://oyabrowser.com/mcp/pool"
http_headers = { Authorization = "Bearer YOUR_API_KEY" }
default_tools_approval_mode = "approve"
```

### claude.ai and Claude Desktop

On a Pro, Max, Team or Enterprise plan you need no API key. Open Settings, then Connectors, then **Add custom connector**, paste `https://oyabrowser.com/mcp/pool`, and sign in to Oya when asked. Pick the project to connect and allow it. The Claude plan pays for the model; Oya is the browser. The connection shows as a credential in the project's settings, where it can be revoked. Any MCP client that signs in with OAuth works the same way.

### Try it

Paste one of these into your agent. Each one proves a different part of the setup.

Works with the desktop app connected or the cloud endpoint configured.

```
Read https://oyabrowser.com/llms.txt, install Oya on this computer and set yourself up to browse the web through it instead of Chrome. Tell me when you need me to click Connect, then open https://news.ycombinator.com in Oya and read me the top three titles.
```

Expected: the agent installs and opens Oya, asks you for the Connect click, then reads the three Hacker News titles from Oya.

Add https://oyabrowser.com/mcp/pool as a connector with no key, then:

```
List every tool the Oya connector offers and what each one does. Do not call any of them yet.

Now call pool_status.
```

Expected: all 32 tools listed without signing in, then a sign-in prompt on pool_status.

```
Using Oya only, sign in to my GitHub notifications page, tell me how many unread notifications I have, then stop and hand the tab back to me.
```

Expected: Oya opens GitHub, the agent reports the unread count, then the tab returns to you.
