2.13.0-beta.20). Features may change before the next stable release.Switch to stable →ak activity
ak activity
Inspect AgentKit's local activity feed, choose a finite snapshot, aggregate, or live stream, and understand its privacy boundary.
Use ak activity to inspect local skill-run evidence without opening the
dashboard or desktop app. The group command is read-only and selects one of
three views over local data.
Usage
ak activityThe group accepts no positional arguments and has no command-specific flags.
Running it without a subcommand prints its help and exits 0.
Choose a subcommand
| Goal | Command | Behavior |
|---|---|---|
| Read a bounded snapshot | ak activity list | Returns recent activity events, newest first. |
| Aggregate skill usage | ak activity stats | Groups privacy-bounded local usage over a lookback window. |
| Follow new events | ak activity tail | Waits at the current end of the log and streams later appends. |
ak activity list --json
ak activity stats --window 7d --json
ak activity tail --jsonLocal data and privacy
The activity log is ~/.agentkit/activity/events.ndjson. AGENTKIT_HOME
changes the ~/.agentkit base. It is an append-only local NDJSON file written
by activity-producing commands such as ak run; the activity commands do not
truncate or apply retention to it.
Events can contain a timestamp, event kind, skill identifier, runtime,
duration, and subprocess exit code. They do not contain prompt text or skill
arguments. stats can additionally read Claude Code session JSONL files to
derive aggregate Skill-tool usage. No activity subcommand sends analytics,
contacts a provider, checks entitlement, or changes the analytics lifecycle.
The disposable analytics index may accelerate stats only when an existing
generation is valid. Otherwise stats reads the authoritative local sources;
it does not create, enable, refresh, or repair the index.
Output and interaction
list and stats produce one human report or one versioned JSON envelope on
stdout. tail --json is different: it emits one kind=activity.event envelope
per line, so consumers must parse it as NDJSON.
These commands never prompt and ignore stdin. --yes has no effect.
--no-interactive only selects plain rendering. See each subcommand page for
its exact output and cancellation behavior.
Exit behavior
| Exit | Meaning |
|---|---|
0 | Help or the selected read completed; cancelling tail through the command context also exits 0. |
1 | A local source read, scan, or output write failed; an unexpected positional argument also currently exits 1. |
2 | A flag was invalid, or stats --window could not be parsed. |
Activity runtime failures in --json mode are not routed through the shared
JSON error renderer. They can therefore return a nonzero status with empty
stdout and stderr. Treat the process status as authoritative and validate a
success envelope before reading it. Flag-parse failures remain plain-text usage
on stderr.
On macOS, the CLI-wide first-launch guard can also write
~/.agentkit/.first-run-darwin and print a one-time hint to stderr before this
command runs.