Overview

K-AI exposes two Model Context Protocol (MCP) servers — one for Retrieval (5 tools, the canonical replacement of the /search endpoint that was removed from the K-AI Instance API in April 2026 — see Removed endpoints) and one for Audit (21 curated tools). MCP clients (Claude Desktop, Cursor, Mistral Le Chat, any MCP-compatible agent) discover OAuth at each MCP URL and authenticate the end-user once. A single token grants access to every K-AI instance the user is entitled to, on both servers. The host LLM does the orchestration; K-AI exposes primitives, not a chat strategy.

Two MCP servers

https://api-retrieval.kai-studio.ai/mcp — 5 tools. The Activate surface over Retrieval. See Retrieval tools.
https://api-audit.kai-studio.ai/mcp — 21 tools. The Activate surface over K-AI Audit, organised conflict-first. See Audit tools. Both servers follow the schema stability policy.

Both servers use Streamable HTTP transport and are stateless (no server-side sessions).

Authentication

Two ways in, one authorization chain.

Interactive clients — OAuth 2.1. Both surfaces share a single Authorization Server: https://auth-api.kai-studio.ai. Clients discover it via /.well-known/oauth-authorization-server on either MCP host. Tokens are signed JWTs and are accepted on both surfaces — register your client once, use it everywhere. See OAuth 2.1 for the full flow.

Autonomous agents — Organization API key. A key created in the K-AI Studio portal is accepted directly in Authorization: Bearer on both MCP endpoints, with no consent screen and no exchange step. This is the path for agent platforms that have no human to authorize anything. See Autonomous agents (API key).

Either way, identity is propagated into every tool call and authorized the same way — group-based RBAC, per-instance visibility, and instructions.

Quick start

The fastest path is Cursor: two one-click install buttons (one per server). Claude Desktop and Mistral Le Chat require UI steps via their respective Custom Connector panels — see Connect a client.