Skip to content
AI observability for agents

See how your AI agents run.

Inspect recorded prompts, model calls, tool use and failures. Follow conversations across runs, compare latency, and understand estimated spend.

Free plan, no card. Node SDK setup in the quickstart.

support-triage · run_8fa21c
Example1 failure
agent.run12.9s
plan0.7s
retrieve_docs1.2s
model.completion2.4s
tool.search_knowledge1.0s
tool.create_ticket0.9s
respond0.5s
Status
Failed
Error
HTTP 500
Duration
12.9s
Cost
$0.0062

Records agent runs from

  • TypeScript
  • Node.js 18+
  • OpenAI
  • Anthropic
  • REST ingest API
Trace every agent run

Understand each step, not just the answer.

  • Execution traces

    Follow instrumented steps in a waterfall, inspect logs and retry events, and open failed runs on the root-cause span.

  • Prompts and tool calls

    Inspect captured prompts, model outputs, tool inputs and results alongside reported token usage and timing.

  • Search and filter runs

    Search trace names, errors and step names. Narrow results by status, duration, cost, environment and time.

  • Sessions and conversations

    Group runs by session and end user, read retained conversation turns, and open the trace behind each turn.

support-triage · run_8fa21c
Example1 failure
agent.run12.9s
plan0.7s
retrieve_docs1.2s
model.completion2.4s
tool.search_knowledge1.0s
tool.create_ticket0.9s
respond0.5s
Cost and performance

Understand your agents’ cost and latency.

  • Recorded cost estimates

    See estimated spend and tokens per run and model call. Unknown prices stay visibly unpriced.

  • Spend breakdownsPro

    Split spend by model, provider, agent and environment over your selected range, then inspect recent expensive runs.

  • Model performancePro

    Compare model usage, cost, failures and latency percentiles, with first-token timing when reported by the model call.

  • Tool reliabilityPro

    Compare tool calls, failures, retry attempts and latency. Follow recent errors into the exact trace step.

Cost analytics
ExampleLast 7 days
Total spend
$18,420
Avg / run
$0.038
Tokens
24.8M

Spend by model

Project visibility

Find failures and slow steps.

  • Project health

    Requests, failure rate, latency and spend for the range you pick, compared with the period before it.

  • Failure triage

    Recent failures list the error and open the run behind it in one click.

  • Cross-trace span explorerPro

    Filter recorded steps by kind, model, tool or duration. Explore latency distributions and jump into matching traces.

  • Usage and retention

    Track your monthly trace allowance and sampling status. See your plan’s retention window and keep aggregate trends after traces expire.

Production health
ExampleLast 24 hours
Traces
12,480
Failures
1.3%
P95 duration
1.84s
How it works

From prompt to production. Tracehatch records the run.

import { init, trace } from "@tracehatch/sdk"

init({ agent: "support" })
await trace("run", agent)

Example · set TRACEHATCH_API_KEY and TRACEHATCH_BASE_URL first.

01Connect

Open project setup and run the guided installer in your existing app. Review the detected framework and file changes, then connect your development environment.

Recorded trace
Example428ms
  • prompt86ms
  • tool call221ms
  • response121ms

02Capture

Trigger an AI action in your app. Supported model calls are captured automatically; setup opens the recorded run so you can check the result.

Latency by model
Examplep95
  • gpt-4.1
  • claude-sonnet
  • gpt-4.1-mini

03Analyze

Inspect traces, find slow steps, break spend down by model and resolve failures.

Production
ExampleLast 24 hours

+18% P95 duration vs previous period

3 failed runs · support-triage

04Monitor

Review failure rate, latency and cost by environment, and compare the selected period with the one before it.

Who it is for

Built for the people who keep agents running.

One trace, read three ways — by the engineer who wrote the agent, the team that runs it, and the lead who pays for it.

AI engineers

Instrument your agent, then inspect failed runs step by step, including captured prompts, tool results and logs.

Platform and SRE teams

Track failure rate, latency and throughput by environment, and investigate changes across time periods.

Engineering leads

Understand estimated spend within each project, with breakdowns by model, provider, agent and environment on Pro.

Integrations

Connect with your existing AI stack.

Capture supported OpenAI and Anthropic calls automatically in Node.js. Other server languages can use the gateway or send model calls to the log endpoint.

OpenAI

04
  • Chat Completions
  • Responses
  • Embeddings
  • Streaming

Anthropic

02
  • Messages
  • Streaming

Node.js SDK

04
  • TypeScript
  • Node.js 18+
  • Custom spans
  • Tool retries

Developer tools

03
  • REST ingest API
  • Project API keys
  • Personal access tokens

Install the versioned Node SDK using the quickstart.

Start with the quickstart
Pricing

Pricing grows with your AI apps.

Start free with traces, sessions, the dashboard and cost totals. Pro and Enterprise include deeper analytics and longer retention.

Starter

Free

For developers exploring AI observability.

  • 10,000 traces per month
  • 3 projects
  • Dashboard, traces and sessions
  • Cost totals and token usage
  • 7-day retention

Pro

Advanced analytics

Previewaccess

For deeper cost and performance analysis.

  • Unlimited metered traces
  • Unlimited projects
  • Cost, model, tool and span analytics
  • 90-day retention

Enterprise

Custom

For projects that need longer trace history.

  • All Pro analytics
  • No plan-based trace expiry
  • Unlimited projects
  • Usage tracking

Plan changes and payments are not available in the app during preview. Contact us about Pro or Enterprise access. Compare every plan.

FAQ

Frequently asked questions.

What engineering teams ask before instrumenting their first agent.

Create a project, open setup and run the guided installer in the server app that makes your AI calls. It detects your framework, connects through your browser and previews the configuration changes. Restart your app and trigger an existing AI action, then open its recorded run. Manual and AI-assisted instructions are also available.

Stop guessing why your agents fail.

Follow recorded runs from model calls to tool results. Find failures, inspect slow steps, and understand estimated spend.