Public audit trail
Every approval my agents asked for since July 4, 2026. Raw args redacted, counts real.
Aadil Ghani, founder of Pushary
49
Total decisions
49%
Auto-cleared by policy
9s
Median approval latency
0% / 100%
Questions vs commands, when asked
One row per decision. Tool name, which agent, what was decided, and how long the human took. AUTO rows never reached the phone, so the questions-vs-commands split is measured over the decisions that did. Commands, file paths, prompts, and question text never make it into this export.
Each row keeps the tool name, the agent, the decision, the date and hour, and the seconds between ask and answer. The export drops everything else: no shell commands, no file paths, no targets, no prompts, no diffs, no question text. The counts and timings are exactly what the audit logs recorded.
Fair objection. Every row here exists because an agent hit a permission gate and Pushary logged what happened next. That is the pitch, not a trick: run your agents through one policy and you get this record as a byproduct. You do not have to trust the summary numbers at the top, you can count the rows yourself.
Exported read-only from the founder's own Pushary workspace on 2026-07-04, covering the previous 30 days. Questions and answers come from the decision log. Auto-cleared rows come from the action receipt log, and any action matching a logged human decision is dropped from the auto count so nothing is counted twice. Regenerated from the audit logs, never edited by hand.
Set one permission policy across Claude Code, Codex, Cursor, and Hermes. Approve the risky calls from your phone. Every decision lands in an audit trail like this one. 7-day free trial.
