Sign-in, elicitation & sampling

OAuth sign-in

Many hosted MCP servers (GitHub, Linear, Notion, Atlassian, Sentry…) use the MCP authorization spec instead of API keys. Moka handles the whole flow:

1. When a server answers 401, the server shows Sign in (in the sidebar and in Settings → MCP servers).
2. Moka discovers the authorization server, registers itself (dynamic client registration) and opens the login page with PKCE.
3. The provider redirects back to http://localhost:<port>/oauth/callback, and Moka connects.

Tokens are stored in ~/.moka/oauth.json (readable only by you) and reused until they expire. Sign out forgets them.

moka.json (optional settings)

{
  "mcpServers": [
    { "id": "linear", "name": "Linear", "transport": "http", "url": "https://mcp.linear.app/mcp" },
    { "id": "corp", "name": "Corp tools", "transport": "http", "url": "https://mcp.corp.example/mcp",
      "oauth": { "clientId": "moka-dev", "clientSecret": "env:CORP_MCP_SECRET", "scopes": ["tools:read", "tools:write"] } },
    { "id": "no-oauth", "name": "Static token", "transport": "http", "url": "https://…",
      "headers": { "Authorization": "Bearer ${TOKEN}" } }
  ]
}

• OAuth is used automatically for http/sse servers without an Authorization header. Set "oauth": false to turn it off.
• Use clientId (and clientSecret) when the server has no dynamic registration, and scopes to request specific scopes.

Caution

The redirect URL includes Moka’s port. If you run Moka behind a proxy or on another host, set MOKA_PUBLIC_URL (e.g. https://moka.internal.example) so providers redirect to the right place.

Elicitation: servers that ask the user

A tool can pause and ask the user for input with elicitation/create. Moka shows a form built from the server’s schema (text, numbers, booleans, single and multi-select enums, email/URL/date formats) and sends back the answer, or decline / cancel. URL-mode elicitation (e.g. “finish checkout in your browser”) opens the link.

Try it with the demo server: “Order me a coffee”. The order_coffee tool asks for your drink, size and milk.

Sampling: servers that borrow your model

With sampling/createMessage, a server asks the client to run a completion. Moka shows the request (system prompt, messages, token limit) and, once you approve, runs it with the active workspace’s model. The call appears in the inspector as llm.request / llm.response for that server.

Try it: “Brainstorm names for a dino café” calls the demo server’s brainstorm tool, which samples your model.

sampling (per server)
(unset)	Ask each time
"auto"	Allow without asking
"deny"	Don’t advertise sampling to this server

Both requests go through the same queue as tool approvals. They’re cancelled if you stop the chat, and appear in the inspector as interaction.request / interaction.resolved.