LiteLLM mcp_servers config for Sume: static_headers and one credential

LiteLLM's proxy can front remote MCP servers, so teams can register Sume's hosted MCP server once and let several apps use it. The LiteLLM MCP page lists mcp_servers keys including url, transport (default http, also sse and stdio), server_id, auth_type, auth_value, static_headers and extra_headers, read 2026-10-03. The question for a Sume setup is which of those carries the credential.

The keys you need

LiteLLM also lists api_key, basic, authorization, token, oauth2 and aws_sigv4 as auth_type values, and oauth2_flow, token_url, client_id, client_secret and scopes for the OAuth path. For Sume, a key sent as a bearer token is the simplest route; the Sume OAuth page covers the interactive alternative.

mcp_servers:
  sume:
    url: https://mcp.sume.com/mcp
    transport: http
    auth_type: bearer_token
    auth_value: os.environ/SUME_API_KEY

One credential, not two

Sume's SDK docs say the REST API rejects a request carrying both Authorization and x-api-key with a 401 "Send only one API key credential." If you use auth_type and also add the same key in static_headers, the proxy may send both. Pick one place for the credential. This post does not state how LiteLLM merges them; test the outgoing headers.

Scope the key to the job

Sume's gates are enforced on the Sume side: idempotency_key is required on paid and write tools, and max_spend_usd is enforced when you provide it. LiteLLM lists MCP cost tracking and guardrails as features, without detail on the page, so do not assume they replace Sume's own limits.

For a shared proxy, give each team a Sume key with only the access it needs. Under OAuth, mcp:read exposes read-only tools and a paid tool returns insufficient_scope; that is a useful default for a team that only needs to look at assets.

• Reference the key from an environment variable.
• Check the outgoing headers once.
• Require idempotency_key in any client that creates jobs.
• Use LiteLLM's x-mcp-servers header to choose which servers a request may use.