Connecting to the MCP servers is free; the docs server needs no key at all. The workspace server authenticates with a Baseten API key, and you pay Baseten for the compute your model deployments and inference calls consume.

Does it support remote/OAuth?

It is remote-only over Streamable HTTP. The workspace server at https://api.baseten.co/mcp authenticates with a Bearer token (BASETEN_MCP_KEY) rather than OAuth, and the docs server at https://docs.baseten.co/mcp is unauthenticated. There is no local stdio package.

Why are no tools listed here?

The workspace server gates tool discovery behind authentication, so its exact tool names are returned only to a connected, keyed client. Baseten documents the capabilities — operating, debugging, and tuning deployments against live workspace state — but we do not list tool names we cannot verify against the running server.

Baseten MCP server

OfficialBasetenConfig last verified Jun 1, 2026

Baseten's official MCP servers give an agent live access to your model deployments and Baseten's docs: deploy, call, and operate models from your editor.

Baseten MCP is Baseten's official integration for working with the model-serving platform from an AI agent. Baseten hosts ML models and Chains in production with autoscaling GPUs, the open-source Truss packaging format, pre-hosted Model APIs, and rolling deployments across environments. The MCP server connects an agent to the live state of your Baseten workspace so it can act on real deployments rather than guessing — the documented value lands on operate, debug, and tune workflows such as promoting a deployment, adjusting autoscaling, and checking model status, where it cuts the wall-clock time and cost of agent runs roughly in half versus working without it.

Baseten ships two remote Streamable HTTP endpoints. The workspace server at https://api.baseten.co/mcp authenticates with a Bearer token (a Baseten API key with management permissions, supplied as BASETEN_MCP_KEY) and reads and acts on your account's deployments. A separate documentation server at https://docs.baseten.co/mcp needs no key and answers questions grounded in Baseten's docs, code snippets, and tutorials. Both are typically wired up together with Baseten's add-mcp installer alongside the companion baseten skill; the Truss CLI is installed separately when you want to push models from local code.

Quick install

Copy-paste configs are provided for all 8 supported clients. Pick your client below.

Add to ~/.claude.json

~/.claude.json

json

{
  "mcpServers": {
    "baseten": {
      "type": "http",
      "url": "https://api.baseten.co/mcp",
      "headers": {
        "Authorization": "Bearer <BASETEN_MCP_KEY>"
      }
    }
  }
}

Or via CLI

bash

claude mcp add --transport http baseten https://api.baseten.co/mcp

Step-by-step guides

Required configuration

BASETEN_MCP_KEYRequired
Baseten API key with management permissions, from app.baseten.co/settings/api_keys. Passed as an Authorization Bearer header to the workspace MCP server. Required.

What you can do with it

Operate live model deployments

Ask the agent to promote a deployment to an environment, tweak autoscaling, or report current status; it reads the live workspace state through the MCP server and acts on the real account instead of relying on stale assumptions.

Ground answers in Baseten's docs

Connect the keyless docs server at https://docs.baseten.co/mcp so the agent can answer Truss, Chains, and Model API questions with citations from the official documentation while it writes or debugs a deployment.

FAQ

Is it free?: Connecting to the MCP servers is free; the docs server needs no key at all. The workspace server authenticates with a Baseten API key, and you pay Baseten for the compute your model deployments and inference calls consume.
Does it support remote/OAuth?: It is remote-only over Streamable HTTP. The workspace server at https://api.baseten.co/mcp authenticates with a Bearer token (BASETEN_MCP_KEY) rather than OAuth, and the docs server at https://docs.baseten.co/mcp is unauthenticated. There is no local stdio package.
Why are no tools listed here?: The workspace server gates tool discovery behind authentication, so its exact tool names are returned only to a connected, keyed client. Baseten documents the capabilities — operating, debugging, and tuning deployments against live workspace state — but we do not list tool names we cannot verify against the running server.

← Browse all ai-ml servers