One log. Every change. Every actor.
Every operation in Odexy is an action. Every change — and every attempt that was refused or failed — writes one row: who did it, the full input, the output or the error, and whether it was marked dangerous. The web app, the AI, MCP and the API all go through the same actions — one write path, one log.
What each row holds.
- Who
- The person, the AI acting for a person, or the system — and the operator, if someone from our side acted inside your workspace.
- What
- The action, with its full input and full output as they happened — not a summary of them.
- How it ended
- Success, or the error it stopped on — refused by a permission, failed validation, not found. Failed attempts are logged too.
- How serious
- Whether the action is marked dangerous — irreversible, destructive, or sending something outside your workspace. The AI is not allowed to run those; a person has to.
- Reads
- What the AI or a background job reads is recorded as well. A person opening a record is not; what they change is.
- What is left out
- Passwords, keys, tokens and private links are masked before a row is written, and a chat row keeps which conversation and how long the message was, not what it said.
Editing history is detectable.
Every day, that day's rows are hashed and chained onto the previous day's link. Change a past row and that day's link no longer matches, nor does any link after it.
Today only we can run that check; the chain is not yet published for you to verify yourself.
Seeing it, keeping it, taking it with you.
- In the product. Admins browse the activity log and open any entry in full, including what the AI did on someone's behalf.
- How long it is kept. At least 90 days on Free, 1 year on Starter, 2 years on Team, 3 years on Business, 5 years on Scale, 7 years on Enterprise.
- Export. Available on Scale and Enterprise. Today we run the export for you rather than from your settings, and the file lists who did what and when — not the full inputs and outputs. There is no streaming feed to a SIEM yet.
The AI is on the same record.
The AI works through the same actions as people, so everything it does is a row like any other — recorded as the AI, against the person it acted for. Filter the log to the AI and you see everything it did. There is no separate AI log to reconcile.
Every change, on the record.
Every action, every actor, every module — one log, chained daily.