It’s not a log.
It’s evidence.
Tracyn keeps a record of everything your AI agents do, asks a person before the risky parts, and turns it all into evidence your auditors accept.
TracynTimeline
Every action your agents take, each chained to the one before it.
| Action | Type | Status | Latency | When |
|---|---|---|---|---|
send_refund support-bot | External | Approved | 1,180 ms | just now |
lookup_order support-bot | Internal | Completed | 240 ms | 1m ago |
bulk_delete_records ops-agent | External | Rejected | 2,410 ms | 3m ago |
update_crm_record sales-agent | External | Completed | 610 ms | 4m ago |
charge_card billing-agent | External | Approved | 890 ms | 6m ago |
summarize_ticket support-bot | Internal | Completed | 1,420 ms | 8m ago |
screen_resume recruiting-agent | Internal | Completed | 3,050 ms | 11m ago |
draft_reply support-bot | External | Error | 1,320 ms | 12m ago |
sync_inventory ops-agent | Internal | Completed | 730 ms | 15m ago |
send_receipt_email billing-agent | External | Completed | 520 ms | 17m ago |
schedule_interview recruiting-agent | External | Completed | 680 ms | 20m ago |
close_ticket support-bot | External | Completed | 310 ms | 22m ago |
How Tracyn works before it acts
Tracyn nudges until the policy is right
Describe what should need approval in plain English. Push back, refine, and apply. It's a conversation, not a one-shot form.
TracynPolicy
Choose which actions need a person to approve them.
Added two rules. Both pause for a person before they run.
+ action_name: "*refund*"+ action_name: charge_card
When it’s risky, Tracyn asks first
A matching action pauses before it runs. Approve, edit, or reject it from the dashboard, email, or Slack.
TracynApprovals
Risky actions wait here until your team decides.
Evidence, not paperwork
Security questionnaires answered from what your agents really did. Every answer cites the events behind it.
TracynSOC2-Questionnaire-2026.csv
How are high-risk actions controlled before they run?
Your audit trail, inside any chat
Tracyn ships an MCP server. Ask what your agents did, clear a pending approval, or draft a questionnaire answer from Claude, ChatGPT, Grok, or any MCP client.
Ask it. Approve it. Done.
Deciding an approval needs a reviewer key, so an agent can never approve its own request.
One approval is waiting: ops-agent wants to run bulk_delete_records (external) with {'count': 500}, requested 41 minutes ago. Approve or reject it?
Rejected, with your note attached. It’s logged to the timeline as your decision.
Type / for commands
Trustworthy by construction
The parts an auditor checks first, built in rather than bolted on.
sha256 3f9a…c21e
sha256 b702…9d44
sha256 e18c…05af
Tamper-evident. Every event stores a SHA-256 hash chained to the one before it. Edit a row and the chain visibly breaks.
lookup_order
Runs freely
send_refund
Needs approval
One policy, two outcomes. An action runs automatically or waits for a person. Nothing in between to misconfigure.
SOC 2, auto-mapped. Controls are backed by what your workspace actually did, not a static template.
Built for the audit
5min
To your first logged action
Install the SDK, wrap one tool call, and it's recording.
SHA-256
Chained event hashes
Each event is hashed with the one before it, so history can't be rewritten quietly.
2
Outcomes per action
It runs on its own, or it waits for a person. Your policy decides which.
Frequently asked questions
Every action your agent takes through the SDK: what it was, when it happened, the model and cost behind it, and whether it ran automatically or needed a human. Nothing is summarized after the fact. It's recorded as it happens.