Sign inGet started

Knows

The full record of everything your agents did.

Every session in one place, titled, charted step by step and priced. Search across all of them, replay any one, and see where the money went.

Get startedSee how it works
Works withClaude CodeCodexpioh-my-piPrimeCursorAmpOpenCodeAntigravityGrokHermesOpenTelemetry · coming
tracez.recurz.ailivesample data
The Tracez session list for an example 25-person team: 1240 sessions in the last 7 days, 9 running now; PRs merged, sessions, sub-agents, tokens, spend and loops caught in the last 24 hours with a heat strip per agent; each session with its PR, open findings, failed steps and a bar chart of its steps; today’s spend, ingest and what is running now on the right.
Titled by a model, every step chartedFindings on every sessionToday’s spend, and where it’s heading

01 · Every run

Every run, from every machine.

Sessions land while they’re still running, titled by a model, with a sparkline of what the agent did, what it cost and what went wrong. Even from CI runners that died hours ago.

  • Search every session at once
  • Sub-agents counted under their parent
tracez.recurz.aisample data
Two live session rows: Migrate checkout to the new pricing table, with its PR, 2 open findings, 7 failed steps and flagged moments made files world-writable and drifted from the task; and Add idempotency keys to payout retries, with its PR on auto-merge, 3 sub-agents and discarded uncommitted work flagged. Each has a bar chart of its steps, tokens and estimated cost.
Findings on every sessionUnsafe edits and drift, flagged

02 · At a glance

A whole run on one line.

Every step is a bar coloured by what the agent did. Sub-agents get their own lane. Loops are hatched red.

tracez.recurz.ai/sessions/S1sample data
The whole-session strip for Fix flaky refund webhook test: 64 steps as coloured bars, five sub-agents in their own lanes (three searches running side by side, then a test runner and a reviewer in parallel before the PR), a loop of four failed test runs (the last three hatched red), then recovery and the end; spend per agent below.
Loop ×4: same test, same failureSub-agents in their own lanes

03 · Turn by turn

Read any run, step by step.

A chapter is everything between two of your prompts. Open one and read every command, edit and test result with its real output.

tracez.recurz.ai/sessions/S1/workbenchsample data
The Tracez workbench for Fix flaky refund webhook test: the whole-session strip on top with five sub-agent lanes; eight passages on the left, one per prompt and one per sub-agent; the first passage in the middle with the owner’s prompt and each step; files changed and the loop finding on the right.
Each prompt starts a chapterWhat the agent actually ran

Where the time and money went

Insights from every session. Insights++ when calls go through Steerz.

InsightsCollectorz + Tracez · every account, every past session

  • Tokens per turn, including cache reads and writes
  • Tool calls and their results
  • Sub-agents and compaction
  • Estimated cost, by kind of work and by session
  • Loops and stalls, with what they cost

Insights++Collectorz + Steerz + Tracez

  • Why the cache hit or missed: breakpoints, TTL, prompt or tool changes
  • The exact system prompt and tool list on every turn
  • Hidden helper calls: titles, compaction, classifiers
  • Time to first token, latency, retries and rate limits
  • Cost from what the provider actually metered

Covers calls that go through Steerz with your own API keys. Sending them to Tracez is coming.

How it works

Kept on disk until it’s confirmed.

The collector never sits between an agent and its model, so it can’t slow one down.

  1. Read. From the files your agents already write.
  2. Hold. Each batch waits on disk.
  3. Confirm. A matching receipt clears it.
  4. Replay. Lose the network, reboot, delete the transcript: nothing is lost or counted twice.

Better together

Tracez is where every other product starts

HandzWatches these sessions live and flags loops, stalls and spend. SteerzRoute calls through it and get Insights++. VerdictzTurns these sessions into evals, so a fix stays fixed.

01 · Knows

What did my agents do?

Five products answer it. Each one is useful alone, and each hands its work to the next verb in the loop.

  1. CollectorzThe free collector on every machine. Reads the transcripts your agents already write.See Collectorz →
  2. TracezEvery session in one place, titled, charted step by step and priced.You are here
  3. MemoryzEvery rule and memory file your agents load, on every machine.Open Memoryz →
  4. ArtifactzEverything your agents made, one lineage per thing.Open Artifactz →
  5. DreamzReads every session overnight and writes the morning report.Open Dreamz →

Start with Tracez

Your agents worked all week. Here’s the receipt.

Get startedEmail the foundersSee the API

Available now: the collector on Mac and Linux, the session list, strip and workbench, findings and the retro. Coming: self-serve sign-up, API search, OpenTelemetry.