Track your LLM costs.Coding agents and API calls.

See estimated spend, tokens, and failures from Claude Code, Codex, and Cursor sessions alongside your SDK requests.

Start free See how setup works

trace-flow.dev/app/agents

Illustrative data

Agent analytics

Where agent work turns into spend

Claude + CodexLast 30 days

Estimated cost over time

$164.20

+8.4%

Aug 04Aug 11Aug 18Sep 01

Tokens processed

83.4M

63% generated

Conversations

148

12 active days

Projected 30d

$182

active-day pace

Per-turn context size

See pressure build before a session runs away

p50 / p90

early turnsconversation depthlate turns

Notable changes

Context above 140k7 sessions

Edit failures11.2%

Cost per review$4.18

Two sources of spend. One account.

Use either path or both. Each one records estimated cost, tokens, and failures you can filter by model and source.

Coding agentsAlpha

Find the sessions worth a closer look

The desktop app reads Claude Code and Codex CLI history, plus Cursor on macOS. It parses sessions on your machine and uploads analytics, not transcripts.

Cost and context by conversation depth
See how context and estimated cost grow through a session.
Tool reliability
Compare failure rates by source, model, or repository.
Review and file attention
Connect agent spend to the code and reviews it touched.

API calls

Every model request, priced

Route your SDK through the Trace Flow gateway. Responses keep streaming while each request records cost, tokens, latency, and errors.

ProviderModelTokensLatencyCost

Anthropicclaude-opus-4-718.2k4.8s$0.42

OpenAIgpt-5.59.7k2.1s$0.18

Googlegemini-2.5-pro22.4k3.6s$0.11

OpenRouterqwen3-coder31.8k6.2s$0.09

Illustrative request data

Set up in a few minutes.

Start with the path that matches how you use models. You can add the other one later.

1
Install and sign in
Download the desktop app for macOS or Windows and connect it to your account.
2
Choose your sources
Review the coding agents it detects, then start syncing.
3
Keep working
It syncs from the menu bar or system tray. Pause it there whenever you like.

Works with Claude Code and Codex CLI on macOS and Windows. Cursor capture is macOS only. Coding-agent analytics is in alpha.

Download the desktop app

What we keep, and what we don't.

What each path sends to Trace Flow, and how long it stays.

Transcripts are parsed on your machine: The desktop app uploads typed facts such as model, tokens, tool outcomes, and redacted excerpts. It does not upload raw transcripts.
Provider keys pass straight through: Your provider API keys authenticate upstream. They are never written to our database, analytics, or stored bodies.
Bodies are encrypted, or skipped: Stored requests and responses use per-organization AES-256-GCM keys. Send X-Trace-Flow-Omit-Body: true to keep usage metadata only.
Retention you can plan around: Request traces are viewable for 7 days on Hobby and 30 on Pro. Coding-agent analytics expire after a year. Monthly model usage totals are kept for five.