Skip to content

Provider Broker

The Provider Broker is the platform's capability gateway: the single place where upstream provider credentials live, and the only way agents reach search, scraping, speech, voice, media, browser automation, and brokered MCP services.

Provider broker — tenants to capability providers
One hop between every tenant and every capability: scoped tokens in, audited, budgeted provider calls out.

Why a broker at all?

Without a broker, every agent that needs web search, voice, or transcription must receive an upstream API key — and every key in a tenant pod is a liability: leak surface, rotation burden, and no way to budget per tenant.

The broker inverts that:

  • Keys live in exactly one place (the broker deployment), mounted from Infisical-managed secrets.
  • Tenants get tokens, not keys. A token grants a named capability with daily request budgets and cost caps.
  • Every call is audited: provider, capability, tenant, units, estimated cost, latency, and status — with no sensitive payload logging.
  • Failover is built in: transcription chains Deepgram → OpenAI; web search chains SearXNG → Serper → Brave → Tavily; scraping chains Crawl4AI → Firecrawl. Each rung can be forced, skipped, or circuit-broken.

Design principles

  1. Fail closed. Missing token, missing key, unknown scope, unregistered service → denied.
  2. Least privilege. Scopes are per capability; quotas are per tenant per day; the token and the gateway slug must match.
  3. Never pass through. The broker is not a generic proxy: only registered services with allowlisted hosts can be reached, and sensitive headers are stripped.
  4. Humans stay in the loop for money. Payment flows run in a live browser the human controls; the broker refuses to carry card data (Luhn-validated rejection).

Endpoints (public shape)

Group Examples
Health & introspection /healthz, /readyz, /v1/capabilities, /v1/services, /v1/limits/status, /v1/circuit/status, /v1/usage/summary
Search & research /v1/search/brave, /v1/search/web (ladder), /v1/search/perplexity
Scraping /v1/scrape
Audio & voice /v1/audio/transcriptions (STT), /v1/tts/speak, /v1/voice/telnyx/call, /v1/voice/telnyx/hangup, /v1/voice/telnyx/webhook (signed), status & metrics
Browser automation /v1/browser/sessions, /v1/browser/browsers, /v1/browser/profiles, /v1/browser/workspaces, self-hosted task/handoff/payment-intent proxies
Media /v1/media/higgsfield/generate
MCP proxy /v1/mcp/<service> — brokered streamable-HTTP MCP servers
Integrations /v1/composio/read (fixed, least-privilege canary operation)

Read next

See Capabilities for the details of each capability and Auth & limits for the token, scope, and quota model.