No. TraxAgent reads usage and cost reports, which contain token counts, model names, API key names and dollar totals. No prompt, output or end-user data is collected or stored.

No. The open-source collector runs on your side, from cron or a GitHub Action, with your admin key in your own secrets. It sends only the report data to TraxAgent with an upload token that can do nothing else.

An upload token can only add snapshots to your organization. An API token can only read your spend and fixes. Neither can change anything at Anthropic or OpenAI. Both can be revoked instantly, and TraxAgent stores only a fingerprint of each, never the token itself.

Snapshots are stored in a managed Postgres database. Ask and we delete every snapshot for your organization.

No. There's no proxy or SDK wrapper, so TraxAgent can't slow down or break your production traffic.

About five minutes: create an account, run the collector command once, and your first audit appears. Scheduling it nightly takes one more step.

Claude through Anthropic's API and OpenAI today. AWS Bedrock, Google Vertex AI and Azure OpenAI are next.

Usage reports can't split a shared key, so TraxAgent flags it. Giving each team or feature its own key, workspace or project fixes that.

Mostly from batching scheduled jobs, prompt caching, the right cache duration, and moving older models to cheaper successors. Each fix shows its monthly range and the steps to ship it.

After a fix ships, TraxAgent compares cost per million tokens on each API key with your baseline and multiplies by your current volume. Traffic growth doesn't count as savings, and newer models' larger token counts are adjusted for.

Fixes that trade quality for cost are labelled 'needs quality test'. Model comparison runs your real prompts on both models and scores them before you switch.

The low end uses cautious assumptions; the high end is the likely result when the fix lands fully. Verified savings replace the range with a measured number.

By your monthly AI API spend, not by seats. Viewers are free, so finance and managers can see the numbers.

Yes, forever: one snapshot at a time and your top 3 fixes in dollars.

Any time from Plan and billing. Upgrades start immediately; cancelling keeps your plan until the end of the period.

Our engineers ship the fixes with your approval. The fee is a share of savings TraxAgent verifies, for 12 months, on top of any plan.

Create a read-only API token, then connect the TraxAgent MCP server in Claude Code or any MCP client, or call the REST API. Answers match the dashboard exactly.

An hourly measurement of time to first token, output speed and cost per task for major Claude and OpenAI models, all on the same task.

No. The comparison runs on your machine with your provider keys. Only scores, timings and costs are uploaded.