Authentication

The Kafkai MCP server supports two authentication methods. Both resolve to your Kafkai account — tools always see your own projects and spend your own credits, whichever method the client uses.

OAuth 2.1

The server implements the standard MCP authorization flow, so compliant clients need nothing beyond the server URL:

• Discovery. Clients find the authorization endpoints through the standard well-known metadata (RFC 8414 and RFC 9728) published under https://kafkai.com/mcp.
• Dynamic client registration. Clients register themselves on first contact. There are no client IDs or secrets to copy.
• PKCE. Authorization uses PKCE with the S256 challenge method.
• Consent. The browser opens a Kafkai page where you sign in with your account and approve the client's access request.
• Token lifetime. Access tokens are short-lived. Clients hold a refresh token and rotate it automatically, so you sign in once, not every hour.
• Revocation. The server supports token revocation; removing the connector in your client ends the session, and short token lifetimes bound how long a stale token stays valid.

Scopes

Two scopes exist:

• kafkai:read — read project and keyword data.
• kafkai:write — create and delete projects, queue research updates, and publish articles. Both scopes are granted by default.

Static API tokens

API tokens are long-lived credentials for clients that cannot open a browser. They are created on the API tokens page and sent as a Bearer header:

Authorization: Bearer YOUR_API_TOKEN

Properties worth knowing:

• The raw token is shown exactly once, at creation. Kafkai stores a SHA-256 hash, so a lost token cannot be recovered — only replaced.
• Tokens do not expire until you revoke them.
• Each token records when it was last used, which helps you spot tokens that are no longer needed — or used unexpectedly.
• Revocation is immediate.

See Connect Hermes and Other Clients for client setup.

What a connected agent can do

Anything you can do with your account's data: read projects, keywords, and articles; create and delete projects; queue research updates. Every data-returning call spends credits from your balance.

Connect only clients you trust, and revoke access you no longer use.

One thing the server never sees: your agent's LLM credentials. The writing happens on your side of the connection, with whatever model your agent runs.