Integrations
Public API
Every page in this section describes one client that connects to your tenant with an API key and nothing else. A page exists only for an integration that was verified end to end against a tenant on the current build: connected, listed what it should list, made a real call, and showed up where the platform records calls. Each page has the same shape: prerequisites, connect, verify, what does not carry over, troubleshooting.
Who can use it: anyone holding an API key. A tool or route needs the scope named on its reference page; the presets on the Developer Hub's API keys tab cover the common bundles. The tenant must be licensed for the public API.
Two ways in
| Path | Endpoint | Speaks | Fits |
|---|---|---|---|
| Public API | https://<your tenant host>/public/v1 | The OpenAI request and response shapes on the model routes, plus the platform's own routes for agents, sessions, search, tools and evaluations | Client libraries and chat front ends that take a base URL and a key |
| MCP server | https://<your tenant host>/mcp | Model Context Protocol over streamable HTTP, JSON-RPC 2.0, nineteen tools | Coding agents, assistants and any MCP client library |
Both accept the same kapi_ key as a bearer token. Both addresses, with your host filled in, sit at the top of Studio's Developer Hub.
Verified integrations
Three kinds of client appear: libraries and a chat front end that take the public API's base URL and a key, MCP clients that take the MCP endpoint and a key, and the OpenAI Agents SDK, which uses both.
| Integration | Path | Verified with | Page |
|---|---|---|---|
| OpenAI Python SDK | Public API | openai 2.54.0 | OpenAI Python SDK |
| OpenAI Node SDK | Public API | openai 7.15.0 | OpenAI Node SDK |
| LangChain | Public API | langchain-openai 1.6.2 | LangChain |
| LlamaIndex | Public API | llama-index-llms-openai-like 0.8.0 | LlamaIndex |
| Vercel AI SDK | Public API | ai 7.0.99 with @ai-sdk/openai-compatible 3.0.48 | Vercel AI SDK |
| LiteLLM | Public API | litellm 1.100.1 | LiteLLM |
| Open WebUI | Public API | Open WebUI 0.11.3 | Open WebUI |
| OpenAI Agents SDK | Both | openai-agents 0.22.2 | OpenAI Agents SDK |
| Claude Code | MCP server | Claude Code 2.1.79 | Claude Code |
| Python MCP SDK | MCP server | mcp 2.2.0 | Python MCP SDK |
| TypeScript MCP SDK | MCP server | @modelcontextprotocol/sdk 1.30.0 | TypeScript MCP SDK |
Pages for OpenCode, Claude Desktop, Cursor, Codex and Visual Studio Code follow as each completes the same verification on the current build. Until then, Connect a client carries the ready-made MCP snippets for Claude Desktop and OpenCode, and Use an OpenAI client library covers the base-URL pattern the libraries share.
How a page verifies
- Connected means the handshake completed: an MCP client lists nineteen tools, or a library lists your tenant's models.
- A real call means one tool call or one completion that returned content. The handshake alone is not proof of a valid key: the MCP server accepts the handshake from any key that starts with
kapi_and refuses only the first tool call. - Recorded means the call left a trace: the key's row on the Developer Hub's API keys tab shows a new last-used time, and a call that reached a model appears in the Control Room's AI Gateway under Logs as a line naming the provider, model, tier, tokens, cost and latency. That line does not name the key or the user.
What never carries over
- The client's own model and billing. Claude Code, Cursor and their kin keep talking to their own model; the MCP server gives them your tenant's tools, not a model.
- Sign-in. There is no OAuth on either endpoint on the current build; the key in a header is the whole of authentication, so keep it out of committed files and use your client's environment-variable expansion where it offers one.
- Session state on the MCP server. It issues no session id, so every call stands alone; conversation memory lives in the platform's sessions, addressed by
session_idon the chat tool.