Trajekt records every step of every run — reasoning, tool calls, inputs and outputs, tokens, latency, cost — so when an agent goes off the rails in production, you can replay exactly what happened.
npm i @trajekt-ai/coreStep through a run like a timeline. Each step shows the prompt, the model's reasoning, the exact tool arguments and what the tool returned.
{
"project": "PAY",
"summary": "NPE in checkout discount",
"priority": "High"
}Tokens, latency and cost are broken down per step. Retries, repeated tool calls and loops are flagged on the timeline, so the expensive part is the first thing you see.
Put two runs side by side, aligned step by step. Added steps, wrong tools and changed outcomes stand out, next to the configuration that changed between them.
Load an example — or your own export — and replay it step by step, with loops, retries and failures flagged.
Langfuse trace export, OTLP/JSON or a Vercel AI SDK result — drop it anywhere on this panel. It is read in your browser by @trajekt-ai/core; nothing is uploaded.
#7 and #9 repeat an earlier call with identical arguments. 2 retries after a failure. The run failed.
{
"query": "checkout/discount.ts",
"scope": "blobs"
}tool path: sentry.getIssue → gitlab.searchCode ×3 → gitlab.readFile → jira.createIssue → gitlab.createMergeRequest → slack.postMessage
Example files: otlp-sentry-triage.json · langfuse-support-bot.json · ai-sdk-trip-planner.json
Point any OpenTelemetry exporter at Trajekt. Spans following the GenAI semantic conventions or the Vercel AI SDK telemetry are understood as-is.
Every run becomes a trajectory: steps, tool arguments and results, tokens, latency and cost, grouped per agent and version.
Open any run on a timeline, compare it with another, and see when a new version starts looping, failing or costing more.
// instrumentation.ts — standard OpenTelemetry, nothing proprietary
import { NodeSDK } from '@opentelemetry/sdk-node';
import { OTLPTraceExporter } from '@opentelemetry/exporter-trace-otlp-http';
new NodeSDK({
serviceName: 'sentry-triage',
traceExporter: new OTLPTraceExporter({
url: 'http://localhost:3000/v1/traces',
headers: { Authorization: `Bearer ${process.env.TRAJEKT_API_KEY}` },
}),
}).start();
// With the Vercel AI SDK, turn telemetry on for each call:
await generateText({
model, prompt, tools,
experimental_telemetry: { isEnabled: true, functionId: 'sentry-triage' },
});Already have traces? @trajekt-ai/core reads them in your own code and tests:
import { normalize, computeMetrics } from '@trajekt-ai/core';
// A Langfuse export, OTLP/JSON spans or an AI SDK result
const run = normalize(trace);
const { status, costUsd, redundantToolCalls, retryCount } = computeMetrics(run);Prompts and tool results often contain customer data. Trajekt is open source and runs with one Docker Compose file, next to your agents.
$ git clone https://github.com/LeoLR/trajekt.git
$ cd trajekt
$ docker compose up -d
✓ Trajekt is running at http://localhost:3000