MCP and browser tools

Lumo does not hard-wire a list of commands into the model prompt. It runs one or more Model Context Protocol servers and lets the agent discover their tools at runtime. The same protocol works for built-in browser tools and remote servers.

Built-in servers#

Transports#

• HTTP Streamable — preferred for remote servers with a request/response shape.
• SSE — supported for servers that stream events over a long-lived connection.
• Built-in servers are registered in-process; they do not leave the extension.

{  "tool": "page.read",  "input": { "tabId": 42 },  "result": {    "title": "Pricing — Northwind",    "sections": 3,    "words": 1284  }}

Extension providers and WebMCP#

A sibling extension can serve MCP over a chrome.runtime.Port. Lumo discovers it with a single L0 probe, checks enabled / authorized / ready, then opens a standard MCP session. Pages that declare WebMCP tools contribute a per-tab surface instead. Both appear in the same registry as remote servers, so the model sees one tool list.

• Remote servers use HTTP Streamable or SSE; extension providers use the runtime port.
• Disabling a source removes its tools on the next step, with no prompt rebuild.
• Tool names are namespaced; a collision never silently overwrites another server.

The security boundary#

• Every tool call is explicit in the transcript; the agent loop can be stopped mid-step.
• External servers are disabled by default and must be added by the user.
• Tool results are treated as untrusted content; page text cannot silently expand permissions.
• The extension requests only the Chrome permissions a feature needs, and documents each one.

Adding a server#

1. 1

   Open MCP settings

   Settings → MCP → Add external server.

2. 2

   Choose a transport

   Provide an HTTP Streamable or SSE endpoint. Authentication headers are stored locally.

3. 3

   Inspect the tools

   Lumo discovers the tool list and shows connection status before you enable it in a conversation.