Skip to main content
api.yambr.com is a LiteLLM deployment that only exposes the Computer Use MCP endpoint publicly. It is not a model gateway — we don’t proxy chat/completions and we don’t resell OpenAI / Anthropic / Google inference.
Bring your own model. Your Yambr API key is for the MCP tool server at https://api.yambr.com/mcp/computer_use. LLM requests go to your provider (OpenAI, Anthropic, a self-hosted LiteLLM, …) with your key. Pipelines that do OpenAI(base_url="https://api.yambr.com").chat.completions.create(...) will fail — that endpoint is closed.

​
The public endpoint

POST https://api.yambr.com/mcp/computer_use
Authorization: Bearer <yambr-key>
JSON-RPC over Streamable HTTP. Same protocol as any MCP server — see MCP API reference. What you get:
  • Computer Use tools: bash_tool, view, create_file, str_replace, sub_agent
  • Per-chat sandbox containers (isolated Docker) — scoped by x-chat-id header
  • Live browser, terminal, and file preview URLs served from cu.yambr.com

​
Three ways to wire it in

​
A. MCP-native client (Claude Desktop, MCP Inspector, Agents SDK)

Point your client directly at the endpoint. Your client’s host application drives the LLM; our server provides the tools. 
{
  "mcpServers": {
    "yambr-computer-use": {
      "url": "https://api.yambr.com/mcp/computer_use",
      "transport": "streamable-http",
      "headers": {
        "Authorization": "Bearer sk-yambr-..."
      }
    }
  }
}
See Claude Desktop integration.

​
B. OpenAI / Anthropic SDK — tool-use with your own model

Call your own provider. Hand the model a tool spec that forwards tool calls to our MCP endpoint. Many frameworks (OpenAI Agents SDK, Mastra, Vercel AI SDK, LangChain, Anthropic MCP support) have native MCP-tool adapters — point them at https://api.yambr.com/mcp/computer_use with your Yambr key. Minimal shape with the OpenAI Agents SDK: 
from agents import Agent, Runner
from agents.mcp import MCPServerStreamableHttp
from openai import OpenAI

# Your own OpenAI key / account — NOT a Yambr key
openai_client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])

yambr_mcp = MCPServerStreamableHttp(
    params={
        "url": "https://api.yambr.com/mcp/computer_use",
        "headers": {"Authorization": f"Bearer {os.environ['YAMBR_API_KEY']}"},
    },
    name="yambr-computer-use",
)

agent = Agent(
    name="computer-user",
    instructions="Use the Computer Use tools to complete the user's task.",
    model="gpt-4o",                  # ← YOUR model, YOUR provider
    mcp_servers=[yambr_mcp],
)

result = await Runner.run(agent, "Create a pitch deck about coffee")
print(result.final_output)
Equivalent patterns exist for Anthropic’s MCP tool support and for LangChain’s MultiServerMCPClient. The rule: model traffic = your provider; tool traffic = api.yambr.com.

​
C. Self-hosted LiteLLM / n8n / Cursor / Open WebUI

Register https://api.yambr.com/mcp/computer_use as an MCP server in your own tool layer:
  • LiteLLM (self-hosted) — add under mcp_servers in config.yaml; see LiteLLM integration.
  • n8n — MCP Tool node; see n8n.
  • Cursor — MCP server entry in settings.json.
  • Open WebUI (self-hosted) — add it alongside your local Computer Use Server as a second MCP tool.
All of these hand the model to your configured LLM provider; Yambr stays a tool provider.

​
Session identity

Every request should set a chat id so each chat gets its own sandbox: 
x-chat-id: <stable-per-conversation>
x-user-email: user@example.com        # optional, for per-user skills
Same x-chat-id across turns = same sandbox (files persist). New id = fresh container.

​
Streaming and long tool calls

Long-running bash_tool / sub_agent calls stream SSE back over the same MCP connection with 15-second heartbeats. If a client closes the socket early the sandbox keeps running — the sub_agent tmux session is reconnectable from chat.yambr.com’s Terminal tab.

​
Troubleshooting

SymptomLikely cause
401 UnauthorizedKey wrong, revoked, missing Bearer prefix. Rotate at app.yambr.com.
404 Not Found on /v1/chat/completionsExpected — api.yambr.com does not expose chat completions. Point your LLM calls at your own provider.
429 Too Many RequestsPer-key max_parallel_requests=5, or you’ve hit max_budget for the 30-day window. Check the dashboard.
x-chat-id ignored, tools fire against a shared sandboxHeader missing or your framework drops custom headers. Most MCP adapters have a headers or extra_headers option — set it there.
Previews show as raw text in Open WebUIUsing stock ghcr.io/open-webui/open-webui — patches aren’t applied. Self-host? See Open WebUI integration.