Skip to content
TRACE_

How it works

One request. Five steps you can read.

TRACE never hands you an answer without the reasoning behind it. Every run follows the same legible path, from raw intent to verified output, and records itself as it goes.

  1. 01

    Intent

    A request arrives. TRACE parses it into a goal and constraints before anything runs.

  2. 02

    Plan

    The goal is decomposed into typed steps with a token and cost estimate you can read.

  3. 03

    Tool calls

    Each step calls a real, permissioned tool. Arguments and results are logged verbatim.

  4. 04

    Verify

    Claims are grounded against retrieved sources. Failed checks retry within budget or escalate.

  5. 05

    Output

    Only verified results ship, with the full trace attached for anyone to audit.

Anatomy of a trace

Every line earns its place.

  • tickTimestamp + cost. Every step shows when it ran and what it spent.
  • callTool arguments. The exact query or API call, never paraphrased.
  • ✓Verification. Grounding score and citations before anything ships.
trace · run_8f21.log
00:02  ▸ call retrieve(doc="soc2-policy") → 6 chunks
00:03  ▸ call draft(answer, cite=true)
00:04  ✓ verify grounded=0.98 citations=3 → pass
00:04  ▸ emit answer (1,204 tokens · $0.004)
Data center infrastructure

Runs the same in a demo and at 3am in production.

Self-host inside your VPC or run on our cloud. The trace is identical, so on-call never has to guess.

Start free Read the docs ▸