OpenRouter

Record the calls that your app makes through OpenRouter, and see their tokens, latency, and cost in AI observability. Use the Agent SDK with the Vercel AI SDK, or post events from any language with HTTP batch ingest.

What you need

An OpenRouter API key in OPENROUTER_API_KEY.
Your Squasher ingest key and project ID in SQUASHER_API_KEY and SQUASHER_PROJECT_ID. See Create an ingest key.

Record calls with the Agent SDK

Install the packages

npm install @squasher-ai/agent ai @openrouter/ai-sdk-provider

Wrap the AI SDK and call a model

src/ai.ts

import * as aiSdk from "ai";
import { withSquasher } from "@squasher-ai/agent";
import { createOpenRouter } from "@openrouter/ai-sdk-provider";

const openrouter = createOpenRouter({ apiKey: process.env.OPENROUTER_API_KEY });

export const ai = withSquasher(aiSdk, {
  apiKey: process.env.SQUASHER_API_KEY!,
  projectId: process.env.SQUASHER_PROJECT_ID!,
  environment: process.env.NODE_ENV,
  context: { workflow: "support-chat" },
});

export async function summarize(sessionId: string, text: string) {
  const result = await ai.with({ sessionId }).generateText({
    model: openrouter("openai/gpt-4.1-nano", { usage: { include: true } }),
    prompt: `Summarize: ${text}`,
  });
  await ai.flush();
  return result.text;
}

usage: { include: true } asks OpenRouter to return the cost of the call. The SDK reads it from providerMetadata.openrouter.usage.cost.

Check the call in Squasher

Select AI observability in the sidebar, then the Requests tab. The provider is openrouter, and the model is the full OpenRouter model name, such as openai/gpt-4.1-nano.

In a long-running server, you do not need ai.flush() after each call. For all SDK options, see the Agent SDK.

Send events with HTTP batch ingest

Without an SDK, post a batch of JSON events. Each event has an id, a type, a timestamp, and a body. The format follows the Langfuse ingestion API.

curl -X POST "https://ingest.squasher.ai/v1/ai/ingest/$SQUASHER_PROJECT_ID" \
  -H "x-squasher-key: $SQUASHER_API_KEY" \
  -H "Content-Type: application/json" \
  --data @- <<'JSON'
{
  "batch": [
    {
      "id": "evt_trace_1",
      "type": "trace-create",
      "timestamp": "2026-09-25T08:00:00.000Z",
      "body": { "id": "trace_1", "name": "support-chat", "sessionId": "chat_1", "userId": "user_42" }
    },
    {
      "id": "evt_gen_1",
      "type": "generation-create",
      "timestamp": "2026-09-25T08:00:01.000Z",
      "body": {
        "id": "gen_1",
        "traceId": "trace_1",
        "name": "reply",
        "model": "openai/gpt-4.1-nano",
        "provider": "openrouter",
        "startTime": "2026-09-25T08:00:01.000Z",
        "endTime": "2026-09-25T08:00:02.400Z",
        "usageDetails": { "input": 920, "output": 255, "total": 1175 },
        "costDetails": { "total": 0.00123 }
      }
    }
  ]
}
JSON

Squasher answers 207 with a successes list when it accepts the batch. The project ID in the URL must be the project of the key.

`type`	Becomes
`trace-create`	The trace. Its `sessionId`, `userId`, and `environment` apply to the other events in the batch with the same trace.
`generation-create`, `generation-update`	A model call.
`span-create`, `span-update`, `event-create`	A step in the trace.
`agent-create`	An agent run.
`tool-create`	A tool call, with its `input` and `output`.

Fields of a generation body:

Field	Description
`id`, `traceId`, `parentObservationId`	IDs that build the trace tree.
`name`, `model`, `provider`	Without `provider`, Squasher takes it from the model prefix.
`startTime`, `endTime`, `completionStartTime`	ISO 8601 times.
`input`, `output`	Any JSON.
`usageDetails`	Token counts: `input`, `output`, `total`, and optional `input_cached_tokens`, `input_cache_read`, `input_cache_write`, `output_reasoning_tokens`. The OpenAI names `prompt_tokens` and `completion_tokens` also work.
`costDetails`	Cost in US dollars. Squasher uses `total`, or the sum of the other values.
`level`, `statusMessage`	`ERROR` marks the call as failed.
`metadata`, `environment`, `version`, `promptName`, `promptVersion`	Kept on the call.

How Squasher counts cost

When the call reports its cost, Squasher uses it. Otherwise, Squasher computes the cost from the token counts and a public price list, which it updates every 24 hours. Models that are not on the list have a cost of 0.

Limits

score-create events are not supported in HTTP batch ingest. Squasher answers 201 for them but drops them. Send scores with the Agent SDK.
A costDetails total of 0 counts as no cost, and Squasher computes the cost instead.
The key goes in the x-squasher-key header. Never put it in the URL.

Get started

Guides

What you need

Record calls with the Agent SDK

Send events with HTTP batch ingest

How Squasher counts cost

Limits

Next steps

AI observability

Agent SDK

​What you need

​Record calls with the Agent SDK

​Send events with HTTP batch ingest

​How Squasher counts cost

​Limits

​Next steps

AI observability

Agent SDK

What you need

Record calls with the Agent SDK

Send events with HTTP batch ingest

How Squasher counts cost

Limits

Next steps