Our GitHub account is temporarily unavailable: links to the code, releases and issues will not open.

Report a bug or talk to us

AI agents

OmniGet as an MCP server

OmniGet can act as an MCP server, so other AI apps on your computer, such as Claude Code, Cursor, VS Code or Codex, can use its tools: start and follow downloads, run your agents as background jobs, read your companies, recall memory and more. Each client gets its own connection with only the permissions you tick, and the server only listens on your own computer.

Where it is

LLM › Configure › MCP, in the Your endpoint half.

How it works

  • The address is http://127.0.0.1:<port>/mcp, where the port is the first free one between 47720 and 47729. The Endpoint row shows the exact address.
  • It only accepts requests from this computer (loopback). Requests from a web page are refused.
  • Every request carries Authorization: Bearer <token>. The token belongs to one connection you create for one client. It is not the browser extension’s Pairing token.
  • The server is Off by default. While it is off, requests to /mcp are refused.
  • It uses the local bridge, the same local service the browser extension uses. If the bridge is turned off in Settings › Integrations › Browser extension, the endpoint cannot answer.
  • OmniGet must be open for clients to use it.

How to use it

The quick way: Connect your AI tools

  1. Open LLM › Configure › MCP.
  2. Turn on MCP server.
  3. Under Connect your AI tools, find your tool (Installed, not connected) and click Connect.

OmniGet creates a connection and writes it into that tool’s own configuration file. The tool may need a restart to see OmniGet. Connect from inside the tool instead adds the omniget-connect skill to your tools, so you can ask them to “connect to OmniGet” and approve the request in OmniGet.

When a tool asks to connect this way, OmniGet shows … wants to connect to OmniGet with a code that the tool also shows. Click Allow only if you just asked it to connect.

The manual way: create a connection

  1. Turn on MCP server.
  2. Under Connect an AI client, type a Client name and tick the Permissions (see below).
  3. Click Create connection.
  4. Pick your client and click Copy configuration (and Copy token for Claude Code). The token is shown only in this session; store it safely.
  5. Paste the configuration into the client (examples below).

Revoke ends a connection at once.

Permissions

Each connection only sees the tools of the scopes you tick. The manual form starts with the download scopes ticked.

Group Scopes
Downloads Inspect media and downloads, Start downloads, Pause, resume and cancel, Read diagnostics and logs, Inspect produced files, Read own history, Access local network media (optional), Transfer produced files from a selected folder
Agents and orchestration Run your agents as background jobs, Read granted agents and owned missions, Pause and cancel owned missions, Propose missions within a local execution grant, Answer permission questions of owned missions
Companies Read the companies of agents, Act in the companies of agents (issues, comments, wakeups)
Memory and library Recall and reflect over what OmniGet remembers about you, Teach OmniGet facts about you, Read your book library (titles, highlights and text)
Telegram Reading chats and files, and downloading or sending (each message you approve)
App and system Observing and controlling the OmniGet window, and the washmybelly scans and plans
Catalog and sign-in Search the agent catalog and install components after I approve, and asking you to sign in (login status only, never cookies)

With every scope ticked, OmniGet offers 300+ tools, grouped as downloads and media, orchestration (agents, missions, jobs and approvals), companies, memory, Telegram, reading, the native app, washmybelly and the Canvas. … tools exposed on the page shows the current count.

Configure your client

The snippets below use port 47720 and <token> as examples. Copy the real ones from Copy configuration.

Claude Code

Run in your project folder. The first line reads the token from a hidden prompt, so it never lands in your shell history:

read -rs OMNIGET_MCP_TOKEN && export OMNIGET_MCP_TOKEN
claude mcp add --transport http --scope project omniget http://127.0.0.1:47720/mcp --header 'Authorization: Bearer ${OMNIGET_MCP_TOKEN}'

Cursor

In ~/.cursor/mcp.json (or the project’s .cursor/mcp.json):

{
  "mcpServers": {
    "omniget": {
      "url": "http://127.0.0.1:47720/mcp",
      "headers": { "Authorization": "Bearer <token>" }
    }
  }
}

VS Code

In your mcp.json:

{
  "servers": {
    "omniget": {
      "type": "http",
      "url": "http://127.0.0.1:47720/mcp",
      "headers": { "Authorization": "Bearer <token>" }
    }
  }
}

Codex

In ~/.codex/config.toml:

[mcp_servers.omniget]
url = "http://127.0.0.1:47720/mcp"
http_headers = { Authorization = "Bearer <token>" }

Copy configuration also has snippets for Goose and Claude Desktop.

Clients that only speak stdio

omniget-mcp is a small stdio proxy for clients that cannot use HTTP, such as Claude Desktop. It reads the address from OMNIGET_MCP_URL and the token from OMNIGET_MCP_TOKEN:

{
  "mcpServers": {
    "omniget": {
      "command": "omniget-mcp",
      "env": {
        "OMNIGET_MCP_URL": "http://127.0.0.1:47720/mcp",
        "OMNIGET_MCP_TOKEN": "<token>"
      }
    }
  }
}

omniget-mcp is not installed with the app; it comes with the command-line package (Command line).

Common problems

“MCP server is disabled”

Turn on MCP server in LLM › Configure › MCP › Your endpoint.

The client says the token is missing or wrong

Each connection has its own token, shown only once. Create a new connection, copy the configuration again, and Revoke the old one.

The port changed

If another program took the port, OmniGet uses the next free one in its range. Check the Endpoint row and update the client’s address.

The client sees fewer tools than expected

It only sees the scopes of its connection. Create a connection with the scopes it needs.