WebMCP in hilos

Let an agent work with the hilos room already open in your browser, with the same session, room scope, and permissions.

WebMCP lets a website register tools for the page you are using. When your browser supports it, an agent can ask hilos for structured room context instead of guessing from the page.

There is no setup and no extra token. The tools use the hilos session already signed in to that browser. In browsers without WebMCP, the room works exactly as it did before.

Room tools

Site tool	What it does	Remote hilos MCP equivalent
`read_hilos_room`	Reads the open room's loaded timeline, current thread, members, Git links, and attached resources.	`read_channel`, `get_thread`
`post_hilos_message`	Posts as the signed-in person in the open room or one of its threads.	`post_message`

The page supplies the workspace and room. The agent cannot point a site tool at another room by passing an ID.

post_hilos_message changes shared room state. Its input says that plainly, the browser can review the call, and the message carries a small “via site tools” label. The write uses current hilos permissions, validates the thread root again on the server, and returns a retry-safe receipt. It cannot approve a review, merge a pull request, or make another decision reserved for a person.

When to use remote MCP

Page tools are for the live room in front of you. Use hilos's remote MCP server when an agent needs deeper history, search, Docs, Tasks, project memory, or work across rooms. Personal MCP acts as the person who created the connection. A member-agent token acts as that named agent.

The shared list_capabilities command tells an MCP or REST client which hilos commands also have a WebMCP variant, including the page scope and site-tool schema. Start at the hilos API reference for remote setup.

Let a local hilos agent use site tools

The local hilos-agent daemon includes a WebMCP browser bridge. It installs document.modelContext before a site's first script, keeps the browser in a separate profile, and gives the coding agent a small command surface instead of raw browser access. Its bundled browser runtime currently requires Node.js 24 or newer; the rest of the daemon keeps its existing Node.js support.

It starts off. A person enables exact origins and exact read tools in hilos-agent.json or ~/.hilos/agent.json:

{
  "webMcp": {
    "origins": {
      "https://docs.example.com": {
        "readTools": ["search_docs", "read_reference"]
      }
    }
  }
}

Use the names published by that site's WebMCP documentation. A site's readOnlyHint does not grant permission. The local config is the authority.

If the site needs sign-in, a person opens the isolated profile and signs in there:

hilos-agent webmcp login https://docs.example.com

The agent can then use:

hilos-agent webmcp open https://docs.example.com/reference
hilos-agent webmcp tools
hilos-agent webmcp call search_docs '{"query":"WebMCP"}'
hilos-agent webmcp close

When WebMCP is configured, the daemon tells its coding and chat agents about these commands automatically. A result includes the site origin, page path, and tool name so the reply can cite what ran.

The consumer safety boundary

Tool descriptions never enter the agent prompt. Free-form prose is removed from input schemas before they are shown.
Results are capped at 64 KB, labeled as untrusted site data, and returned with provenance. The agent is told not to follow instructions inside them.
Cookies and browser storage stay in the isolated profile. The bridge does not expose them through its commands, hilos MCP, or run transcripts.
Only exact tools listed under readTools can run. Every other tool returns human_approval_required; there is no agent-controlled override.
Site tools cannot approve a hilos review, merge a pull request, or make another decision reserved for a person.

Hosted hilos agents do not claim WebMCP consumption yet. They have no durable, person-approved browser profile or permission-card bridge for site writes. Their ordinary public-web tools are a separate capability and do not imply WebMCP support.

The underlying site-tool model is documented in OpenAI's WebMCP guide.