Skip to content
Hydra

The audit log

Hydra writes one audit log for everything people and agents do on it. It answers the question an owner gets asked after something goes wrong: who asked for this, when, and what did the agent actually do? It lives on your server, next to the rest of Hydra's state, and nobody else holds a copy.

Open it in System → Audit. The owner and admins can read it; other roles cannot.

What is recorded

Each line has a time, the person (the actor), their role, an action name and the details of that action. For example:

Area What a line says
Sign-in Who signed in or was refused, and how.
Chat Each prompt as the person wrote it, with the model that answered and the files attached.
Tools Each tool the agent called for a person: which integration, which tool, whether it wrote, whether it was approved, whether it worked.
Terminals The lines operators type into terminals. A line typed at a password prompt is withheld: the record says it was hidden, not what it was.
ERP Each change-guard review, request, decision and result, with document types, names and counts.
People Invitations, role changes and status changes, with who made them.
Devices Pairing, revoking, and the host name of each page a device read for a chat.
Pictures and jobs Pictures made, edited, saved or deleted; background jobs started, stopped and resumed.
AI API Each call: the key, the model, the time taken and the token counts. Never the prompt or the answer.
View as Start, stop and every refused write, with the owner as the actor.
Settings Changes to models, limits, integrations and approval rules.

The log never holds a secret: no passwords, tokens, API keys or ERP field values.

The hash chain

Every line carries two extra fields: prev, the hash of the line before it, and hash, a SHA-256 of prev plus the line's own content. The first line starts from a fixed value. The log is split into one file per month, and the chain runs across the files.

Change one character in an old line, delete a line or insert one, and every hash after it stops matching. Each time you open the page, Hydra walks the whole chain and shows the result at the top:

  • Verified: N events, chain intact.
  • or Chain broken at file line n, with the reason and how many events before it still verify.

The chain makes tampering visible; it does not make it impossible. Someone with full control of the server could rewrite the chain from a point onward. Keep your backups of the state directory somewhere else, and an old backup will show the difference.

Filters

Narrow the list with:

  • Filter by person: one actor, or Everyone.
  • Filter by action: one kind of event, or Every action.
  • Filter by project: one project, or Every project.
  • From date and To date: a whole-day range.

The newest events come first. Load more fetches the next page. Show event opens the full line, with every field.

For example, to answer "what did the agent change in the ERP for Tomas last week", pick Tomas under person, an ERP action, and last week's dates.

CSV export

The CSV button downloads the events that match your current filters, as hydra-audit.csv, up to 100,000 rows. The columns are ts, actor, role, action, project and detail (the rest of the line as JSON).

A cell that starts with =, +, - or @ is prefixed with an apostrophe, so a prompt can never run as a spreadsheet formula when you open the file.

Where it lives

On a self-hosted server the log is in the state directory (HYDRA_STATE_DIR), under audit/, one YYYY-MM.jsonl file per month, readable only by the Hydra service. A backup of that directory is a backup of the log. See Configuration.

Related