New TypeScript SDK v0.4.0 · OpenAI, Anthropic and LangGraph

Watch your agent think.

Tracify turns every agent run into a map you can read. Each model call, tool call, retry and dollar, wired together so you see exactly where it went sideways.

Start tracing free Explore the demo

Why it matters

Agentsdon'tcrash.Theywander.Asearchcomesbackempty,afallbackmodelguesses,atooltimesout,andtheanswerstilllooksfine.Tracifyshowsthewholepath.

Before / after

Same run. Two ways to read it.

Scroll, or drag the handle. On the left, the run as a wall of logs. On the right, the same run as Tracify draws it, with the failing step lit up.

The run drawn as a clean flow graph with the failed step highlighted
The same run as a dense wall of log lines
Plain logsTracify

How it works

From black box to blueprint.

01 / 03

Capture everything.

Wrap your agent once. Every step streams in as it happens: model calls, tool calls, retrievals, with inputs, outputs, tokens and timing. Secrets get stripped before anything leaves your app.

Hundreds of steps, one clean trace
02 / 03

Follow the money.

Cost is tracked on every step, not just per month. See the one retry that quietly tripled the bill, and sort every run by what it cost.

The step that burned the budget
03 / 03

Find the first crack.

Failures don't start where they surface. Tracify points at the first step that went wrong, so you fix the empty search, not the tool that timed out three steps later.

Root cause, not the symptom

Replaying run · scroll to play

3.4 seconds. One wrong guess.

Time0.00sCost$0.000
0s1s2s3s3.4s

agent · support-agent

llm · plan

tool · orders.lookup

tool · policy.search

0 results

llm · fallback

Root cause · guessed the refund

tool · refund.issue

skipped

llm · reply

Found it: the policy search came back empty, so the agent guessed.Every run, read like this.

Every run, drawn as a map.

Pick a run and Tracify draws it: who called whom, what came back, how long it took and what it cost. The red flag is where you start.

tracify.tech / support-bot / tracing / run_9f3a
AGENTsupport-agentLLMchatTOOLdocs.searchTOOLorder.lookupLLMchat · retryTOOL · FAILEDrefund.issue

Watch one run

See it happen, as it happens.

Every step lands in Tracify the moment it finishes. Here a support agent answers a refund request, and you can see exactly where it goes wrong.

Live · support-agent · run_9f3a

Customer

Hi, can I get a refund on order 4821? It arrived damaged.

What the agent is doing

  1. plan…

Time

0.00s

Tokens

0

Cost

$0.000

plan
orders.lookup
policy.search
fallback
reply

Watching for the first bad step…

Anatomy of a step

Every step, taken apart.

  • 01

    Input

    The exact prompt or arguments that went in, secrets stripped.

  • 02

    Output

    What came back: the reply, the tool result, or the error.

  • 03

    Model + tokens

    Which model answered and how many tokens it used each way.

One step exploded into stacked layers, with the cost layer highlighted
  • 04

    Cost + latency

    What the step cost and how long it took, to the millisecond.

  • 05

    Parent

    The step that called it, so loops and branches stay readable.

  • What the SDK records

The toolkit

Built for the 2am incident.

Plugs into your stack.

TypeScript and Python SDKs, OpenTelemetry, or plain HTTP. Works next to OpenAI, Anthropic, the Vercel AI SDK, LangChain and LlamaIndex.

See integrations
agent.tsAbout ten lines

Search like you think.

Filter by status, cost, latency or tokens. Exclude with a minus.

⌕status:error
status:errorcost>0.05-env:devlatency>3stokens>=10kname:support-agent
01

Secrets stay home

Bearer tokens, API keys, JWTs and passwords are stripped from inputs and outputs before they're sent.

02

Loops make sense

Nested steps keep their parent, so a five-turn loop reads as five turns, not fifty lines.

03

Runs never mix

Concurrent runs stay in separate traces, even when they share a process.

Use cases

Whatever your agent does.

01 / 04

Support agents

Find the decision behind every escalation.

Explore
02 / 04

Research agents

Show the path from question to source.

Explore
03 / 04

Automation agents

Debug the chain, not the final error.

Explore
04 / 04

Tool-calling agents

See what the model asked the world to do.

Explore
Not sure yet?Try the live demo →

Pricing

Free to start. Fair as you grow.

Free

Try it on a real agent.

$0/ mo

  • 25k spans / month
  • 7 days retention
  • 1 member · 1 project
  • Trace viewer
  • Cost visibility
  • Community support
Start free

Pro

Popular

For a builder shipping agents.

$19/ mo

  • 500k spans / month
  • 30 days retention
  • 5 members · 3 projects
  • Evaluations + datasets
  • Prompt versions
  • Slack alerts
Get Pro

Team

For teams running agents in prod.

$39/ mo

  • 2m spans / month
  • 90 days retention
  • 20 members · Unlimited projects
  • Release gates
  • Collaboration + reviews
  • Priority support
Get Team

Enterprise

For larger rollouts.

Custom

  • Unlimited spans, custom retention
  • SSO (SAML / OIDC)
  • On-prem deployment
  • Rollout support
Talk to us

No card needed for Free. Compare every plan →

What is a span?

One recorded step in a run: a model call, a tool call, a retrieval. Plans are sized by spans per month.

Do I have to rewrite my agent?

No. Wrap the entry point once and mark the steps you care about. Everything else keeps working the way it does today.

What can I send traces from?

The TypeScript and Python SDKs, OpenTelemetry, or plain HTTP.

Can I try it without paying?

Yes. The Free plan includes the trace viewer and cost visibility, no card needed.

Stop guessing. Start seeing.

Wrap your agent, run it, open the map. The free plan needs no card.

Start tracing free Read the docs