Gumloop custom MCP server: use a bearer token for Sume

Gumloop's Anthropic models do not forward custom headers, so connect Sume's hosted MCP with a bearer token. Or pick another model for x-api-key.

4 min readSume
All posts

Connect Sume to Gumloop with a bearer token. Gumloop's docs state that Anthropic models do not forward custom headers, so an x-api-key header would be dropped on those models. Sume accepts Authorization: Bearer $SUME_API_KEY at https://mcp.sume.com/mcp, which fits Gumloop's Access Token field.

What Gumloop supports

Gumloop custom MCP (read 2026-10-02)
TopicFact
Public URLHTTPS only, HTTP not supported
Private serversVia tunnel, Enterprise only
Auth optionsBearer token, custom header, OAuth discovery
Anthropic modelsDo not forward custom headers
Alternative modelsOpenAI, Gemini or Groq per Gumloop's note

Pick the right auth

Use the bearer option and paste a Sume API key. That route works for any model. If you prefer a custom header, switch to a model Gumloop says forwards headers. OAuth discovery is also supported by Gumloop; Sume's MCP host offers OAuth with PKCE and consent, mcp:read required and mcp:write opt-in.

Make the flow safe

Give each paid call an idempotency_key. Gumloop flows can retry, and the same key with the same request returns the original job. Optional dry_run and max_spend_usd let a flow check cost before spending.

  • Run tools_list once to see current tool names.
  • Under read-only OAuth, mutating tools return insufficient_scope.
  • Use jobs_wait with up to 20 job_ids; each call holds at most 55 seconds.

Timeouts

Gumloop mentions timeout handling for slow servers. Sume's tools return a job id quickly, and the wait happens in jobs_wait slices, so an individual tool call stays short. If a slice expires with wait_slice_expired, retry with the same ids and never recreate the job.

Sources

Related posts

More in Integrations

All Integrations posts

Written by Sume