MCP

Unbrowse is a remote MCP server: https://unbrowse.ai/mcp (Streamable HTTP, protocol 2025-11-25). Sign-in is OAuth 2.1 with PKCE and dynamic client registration; your MCP client runs it. An API key works as Authorization: Bearer ub_live_….

claude mcp add --transport http unbrowse https://unbrowse.ai/mcp
codex mcp add unbrowse --url https://unbrowse.ai/mcp && codex mcp login unbrowse

{ "mcpServers": { "unbrowse": { "url": "https://unbrowse.ai/mcp" } } }

Agents connect to the hosted server. For hosts that need a local stdio server, or reject tool names with dots (Grok Build), npx -y unbrowse mcp proxies to it with the CLI's sign-in; tool names there use _ (unbrowse_scrape). See cli.md. Shells and scripts use the CLI.

Install the agent skill for operating guidance. Exported core input schemas: tools.json. The connected server's tools/list also includes user-specific tools.

Core tools

Tool	Use
unbrowse_discover	Search your private space, then the public registry
unbrowse_sites	Per-site state: public or behind sign-in, kept session, last sign-in, saved login, learned tools, bot checks
unbrowse_usage	Verified calls this month, rendered runs, passthrough cost, quota left
unbrowse_run	Start a run by capability id or plain-language task; input keyed by the capability's inputs
unbrowse_inspect	Status, requirements, verified result, known effects, routing explanation
unbrowse_resume	{ runId, answers: { <field>: value } } on the same run
unbrowse_cancel	Stop new dispatches; effect receipt
unbrowse_forget	Delete your learned capability, or unpin a public one
unbrowse_browse_open / _snapshot / _act / _finish / _close	Recorded cloud browser
unbrowse_learn	Compile HAR files or traces into a learned.* capability
unbrowse_credentials_list / _request / _status	Password manager: masked hints, save-login links
unbrowse_connect	Connect an app's own MCP server (Linear, Notion, GitHub…) or any remote MCP by url; sign-in and key servers return the link the person opens
unbrowse_import_mcps	{servers: [{name, url} | {name, command, args}], dryRun?}: bring the MCP servers the agent already has into Unbrowse (remote by address, local as the hosted app; never keys)

When three or fewer capabilities match, dedicated unbrowse_skill_* tools are listed with slot schemas from the harness. Your compiled tools appear as my__<site>__<op>; tools/list with a query adds matching public ones.

Per-site servers

Each compiled site is its own MCP server with its tools plus run_task, browse_*, run_status and run_answer:

https://unbrowse.ai/api/v1/sites/<host>/mcp

unbrowse site <host> lists the site's tools.

Sign-in elicitation

When a tool needs a login nobody saved, clients that declared URL elicitation at initialize get JSON-RPC error -32042 with data.elicitations[0].url: a one-time Unbrowse page where the person saves the login. Other clients get the link in the result (signIn.url, or details.url on an autofill error).

Operating order

1. unbrowse_discover (and unbrowse_sites for a site you will sign in to).

2. unbrowse_run with a capability id or task. Pass every input you know.

3. input_required → unbrowse_resume on the same run.

4. no_capability → do it once with unbrowse_browse_*; it is learned.

The full agent contract is skill/SKILL.md.

Live canvas

When listed, unbrowse_canvas_read and _put read or create notes, plans, results and reply drafts on your canvas. Use the same account in both places. Card creation does not send external messages. Updates use revisions and preserve human edits.

unbrowse_scrape reads a page; unbrowse_map discovers same-site URLs. Consult the returned schemas for supported options.