Activity

What is recorded

Activity holds two things: every call an agent made through Inline, and every change to who can call what. A record is written whatever the outcome — calls that succeeded, calls that were refused, calls a rule held, calls that failed upstream, and calls that were never sent because a limit or a brake stopped them.

Owners, admins and auditors read the whole workspace’s trail. Everybody else reads their own: the server answers them with their own calls, and the page labels the tabs accordingly.

The tabs are Tool calls, Script runs and Searches, plus Admin changes and Key activity for those who may read the workspace’s logs. Searches have their own tab because an agent looking for tools is a different question from an agent using one: each search records what was asked and which tools were shown.

Filtering

Calls can be filtered by person, server, tool, outcome, and how the call was made — MCP, HTTP, in a script, or in a registered tool. The outcomes are the ones the rest of the product uses: ok, refused, failed upstream, practice (nothing was sent), not sent because the same call kept failing, not sent because the call allowance ran out, and waiting for approval. A held call can be filtered either as still waiting or as held at any point, decided or not.

New rows arrive at the top of the list while you read; if you have scrolled away, a pill says how many came in rather than moving the page under you. The Live sessions panel shows who is connected now and which sources their session can reach.

What one record holds

Opening a call shows when it happened, the person, the key it acted through, the server, the source, how it was made, how long it took, and the reason if it was not ok. A held call carries what became of it: who decided, when, any note, and a link to the call that ran afterwards. A call that ran after an approval links back to the held one.

Arguments are kept by default, with values that look like credentials removed and the whole set capped at 8 KB. A SHA-256 of the complete set is kept too, so a truncated copy can still be matched against the original.

Responses are not kept. What is always recorded is the response’s size and its SHA-256 — enough to prove two calls returned the same thing, or that a payload somebody shows you later is the one that was served. A source can be switched to keep responses as well, redacted and capped at 16 KB, while it is being investigated; that setting is per source, and the intention is that it goes back off afterwards. A source can also be set to keep no arguments at all, in which case the trail still records who called what, when, and how it ended.

Admin changes record what changed, who changed it, the action, and the details. Secrets never appear: keys are recorded by label and prefix only.

How long it is kept

The two trails on this page — calls and admin changes — are each written with an expiry date. It is 365 days from the record on every plan; plans do not differ in how much history they keep. The deployment can set a different number. Deletion happens through the database’s own expiry policy, so a deployment that has not switched that on keeps records until they are removed by hand.

Not everything in a workspace expires. The source of every script an agent runs is kept in its run record, with the tools it called and how it ended, and those records have no expiry: they stay until the workspace is erased, even after the tool they ran is deleted. So do saved tools and their earlier versions. Requests for keys and one older key log have none either. The privacy policy lists what is kept and for how long.

An approval record — the held call’s own row, with its encrypted arguments — is kept for 30 days after it is decided by default. After that, Activity can still say a call was held, but the approval itself is gone.

Each record is also printed as a single JSON line to the service’s standard output, so a deployment can route the trail to its own logging or SIEM without giving anyone database access. On our own hosted service those lines are also in Google Cloud Logging, where they age out after 30 days and cannot be deleted one by one.

What it does not do

Activity does not record everything. When an agent host asks a connected service for one of its app panels, Inline forwards that request with your credential, and no call record is written for it.

Activity is a record, not an alarm. It does not notify anyone; a call that needs a person is a matter for rules and approvals, and a source or server that is misconfigured is what Checkup reports. What Inline claims and does not claim about its own security posture is set out on the security page.