> ## Documentation Index
> Fetch the complete documentation index at: https://docs.yambr.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Access model — MCP only

> Yambr publishes one public surface: the Computer Use MCP endpoint. Model inference stays on your own provider.

## The one-line version

**Yambr's public API is MCP only.** You get a sandbox, tools, and file hosting. You bring the model.

## What's behind api.yambr.com

| Path                        | Public? | What it is                                                  |
| --------------------------- | ------- | ----------------------------------------------------------- |
| `POST /mcp/computer_use`    | **Yes** | MCP over Streamable HTTP. Authenticate with your Yambr key. |
| `GET /mcp-info`             | Yes     | Tool list, required headers, endpoint URL.                  |
| `POST /v1/chat/completions` | **No**  | Closed. Not a resell endpoint.                              |
| `POST /v1/completions`      | **No**  | Closed.                                                     |
| `POST /v1/embeddings`       | **No**  | Closed.                                                     |
| `GET /v1/models`            | **No**  | Closed.                                                     |

Everything under `/v1/*` is internal plumbing — it's there because `api.yambr.com` is a LiteLLM deployment under the hood, but we don't publish those paths to end users.

## Why MCP-only

A Yambr key costs us compute (Docker sandboxes, disk for artifacts, bandwidth for browser streaming) but **not** tokens. Tokens are paid by you to the model provider you pick.

This keeps two things clean:

* **Billing is local to each provider.** OpenAI bills you for OpenAI usage; we bill you (eventually) for sandbox time. No opaque markup from us reselling inference.
* **Model choice stays yours.** New model drops? Use it the day it ships, without waiting for us to add it to a gateway. Fine-tunes, private endpoints, self-hosted vLLM — all work the same way.

## How "bring your own model" actually looks

Concretely, you have two kinds of LLM traffic in your app:

1. **Model calls** — `client.chat.completions.create(...)` — these go to *your* provider with *your* key. The provider can be OpenAI, Anthropic, Google, a self-hosted LiteLLM, a fine-tune endpoint, anything.

2. **Tool calls** — when the model wants to run `bash_tool`, `view`, `create_file`, `str_replace`, or `sub_agent`, the framework sends that to Yambr's MCP endpoint with your Yambr key.

Every mature agent framework has a clean split for this. Examples:

* **Claude Desktop** — model is your Anthropic subscription; Yambr goes under `mcpServers` in `claude_desktop_config.json`.
* **OpenAI Agents SDK** — model is your OpenAI key on the Agent; Yambr goes in `mcp_servers=[MCPServerStreamableHttp(...)]`.
* **LangChain / LangGraph** — model is your chat model binding; Yambr is an MCP tool via `MultiServerMCPClient`.
* **LiteLLM (self-hosted)** — model routes you already have; Yambr added under `mcp_servers` in `config.yaml`.
* **n8n** — model is the AI Agent node's credential; Yambr is the MCP Tool node.
* **Cursor** — your Cursor subscription runs the model; Yambr is an MCP server in `settings.json`.

## FAQ

**"So I still need an OpenAI / Anthropic key?"**
Yes — or any OpenAI-compatible key. The Yambr key does not give model access.

**"What about `chat.yambr.com`? Isn't that running models?"**
Yes — `chat.yambr.com` is a hosted Open WebUI with models we pay for, as a convenience for browser users. That's not exposed as a public API.

**"Can I use sub\_agent (Claude Code) without an Anthropic key?"**
The `sub_agent` tool launches Claude Code inside the sandbox. It needs an `ANTHROPIC_AUTH_TOKEN` — in managed Yambr we supply that for you inside the sandbox env for the hosted Claude Code runs; self-hosters set it in `.env`. See [Claude Code gateway](/install/configuration#gateway-deploy-routing-claude-through-a-proxy).

**"Will you ever open `/v1/chat/completions`?"**
No plan to. The MCP-only surface is the product.

## Related

* [LiteLLM gateway](/platform/litellm) — the MCP endpoint in detail
* [API keys](/platform/api-keys) — get a key, use it correctly
* [cu.yambr.com](/platform/cu-endpoint) — the artifact host