Integrations
Works with the stack you already run.
PeerLM reads your production traffic from the tracing tool you already have, and sends the verdict back to where your team already works. No app rewrite, and no new dashboard anyone has to remember to open.
Traffic in
A Monitor needs real prompts and the responses your production model actually gave. Connect whichever of these already sees them.
OpenTelemetryRecommended
Point any OTLP exporter at PeerLM — protobuf or JSON, gzipped or not. Works out of the box with the Vercel AI SDK, OpenLLMetry, and OpenInference. This is the highest-fidelity path: it is the only one that preserves span structure and trace ids end to end, which is what lets PeerLM tell an agent step from a single-turn generation and write verdicts back onto the exact trace.
Langfuse
Connect a public/secret key pair and PeerLM pulls observations on a schedule. Cloud or self-hosted. Trace ids survive the round trip, so verdicts can be written back as Langfuse scores.
Braintrust
Connect an API key and a project id and PeerLM pulls project logs. Verdicts go back into the same project as scores and metadata on the original rows.
LangSmith
Connect an API key and a project name. PeerLM filters to llm runs server-side, so chain and tool spans never distort the corpus statistics.
Helicone
Connect an API key and PeerLM pulls request logs on a schedule.
Cloudflare AI Gateway
Point PeerLM at an exported log file from your gateway.
Direct API & file upload
POST batches to /api/v1/ingest from anywhere, or import a CSV, JSON, or JSONL file. A file import carries no trace ids, so verdicts from it cannot be scored back to a source tool.
Verdicts out
A decision nobody sees is not a decision. Every completed Run sends the same verdict — the call, the receipt, and a link to the evidence — to each destination you connect.
Slack
Every completed Run posts the verdict, the parity and savings receipt, and the weakest slice of your traffic, with a direct link to the decision. The call arrives in the channel, not a nudge to go find it.
Signed webhooks
POST the verdict to any endpoint you control. Every delivery is HMAC-SHA256 signed over the timestamp and body, so a captured payload cannot be replayed and a forged one cannot claim to be us. Wire it to PagerDuty, Linear, or your own service.
Scores on the original trace
Write peerlm.verdict and peerlm.parity straight back onto the traces in Langfuse or Braintrust. Your team sees PeerLM's judgment inside the tool they already have open, next to the request that produced it.
Email digest
A recurring digest that leads with the decision — how many monitors are asking for one — rather than a report that work happened.
Routing export
When a Run says route rather than switch, export the recommended per-category routing as JSON, YAML, a LiteLLM proxy config, a Portkey gateway config, or a code snippet.
Connect a source, get a verdict.
Start with one sponsored comparison on your own prompts. No contract.