DocsAdaptersBrowse

Adapters

The SDK captures agents you wrote. Adapters capture the rest: model traffic behind a gateway, framework spans, runs inside the OpenAI Agents SDK, and finished calls a voice platform reports by webhook. Each supplements the SDK; none replaces it: a gateway sees only what crosses it.

  • OpenAI
    Every model call, and handoffs between agents
  • Anthropic
    Every model call
  • LangChain
    Chains, tools and retrievals as they run
  • LL
    LiteLLM
    Everything behind your gateway, including agents you did not write
  • Vapi
    Finished calls, and proof the AI disclosure played
  • OpenTelemetry
    Whatever you already trace
  • Amazon S3
    Write-once storage, where not even we can delete

LiteLLM gateway

One callback line captures every model call from every agent behind the gateway (including agents whose source you don't control), with MCP tool calls promoted to first-class tool_call events rather than buried in metadata.

config.yaml
callbacks: auditant.gateway.handler
stamp the join key on requests
curl $GATEWAY/v1/chat/completions \
  -H "Authorization: Bearer sk-agent-loan-underwriter" \
  -H "x-auditant-agent: loan-underwriter" \
  -H "x-auditant-session: sess_b0f9c646" \
  -d '{"model":"gpt-4.1","messages":[…]}'

OpenAI Agents SDK

Nothing to add: when the agents package is importable, init() registers Auditant on the SDK's trace-processor bus and says so in handle.activated. The manual form remains for hosts that wire things themselves. It is idempotent, so calling both never chains a span twice.

python
import auditant
handle = auditant.init(api_key="ak_…", log_id="t_…/prod", agent_id="support-bot")
# "openai-agents-trace-processor" in handle.activated

# or, with init(instrument=False):
from auditant.openai_agents import install
install(handle)
what maps to what
generation / response  →  model_call        tokens observed
function               →  tool_call         name + input/output hashes
handoff                →  delegation        which agent handed to which
guardrail (triggered)  →  policy_decision   blocked, on the chain

Handoffs matter more than they look: multi-agent delegation is the case supervisors name as hardest to reconstruct, and here it is a first-class chained event.

Voice & vendor webhooks

Voice platforms report finished calls by webhook; the ingest endpoint maps them to lifecycle, per-turn decisions, tool calls, and a first-class disclosure event recording whether the “you are talking to an AI” line was actually played on this call, which nobody else produces as provable record.

direct Vapi connection
POST /v1/ingest/voice/vapi?log=t_…%2Fprod
x-vapi-secret: <AUDITANT_WEBHOOK_SECRET>
Retell or another audio loop, after signature verification
POST /v1/ingest/voice/retell?log=t_…%2Fprod
x-auditant-webhook-secret: <AUDITANT_WEBHOOK_SECRET>

{ "callId": "call_…", "agentId": "receptionist",
  "transcript": […], "toolCalls": […], "costs": {…} }

Vapi connects directly. Retell signs its own vendor-shaped webhook, so a small relay verifies x-retell-signature before mapping the call to the canonical body above; the same route works for any audio loop. The Auditant webhook secret is deliberately not the API key: it can write events and do nothing else: leaking it costs spam, never disclosure.