Automations

An Automation is a saved Lucius conversation that starts itself instead of you typing — the exact same turn as chat, same tools, same permissions, just triggered instead of typed. Nothing behaves differently because nobody’s watching, except the one setting below. (Internally still called “Agent” in tool names and some API fields; everywhere in the product, it’s an Automation.)

Creating one

From the Automations tab, set what it should do (a plain-language instruction), a trigger, and an approval mode. Creating, editing, deleting, or manually running an Automation is itself always gated — it can’t use its own authority to change what it, or another Automation, is allowed to do.

Triggers

Trigger	Fires when
Manual	You (or Lucius, at your request) explicitly start it
Scheduled	Hourly, Daily, Weekdays, Weekly, Monthly, Once on a date, or a custom recurrence rule (ask Lucius to help set an unusual cadence)
Event-triggered	A transaction posts, a balance crosses a threshold, a commitment becomes due, a connector sync fails, or a price moves

Approval mode — the one setting that changes behavior

Each Automation has its own approval mode, independent of your global tool permissions:

Mode	What runs
Manually approve	Pauses on every call that would otherwise ask, and shows the run’s summary before anything applies
Automatically approve	Runs through without a card — except two especially risky operations (below), which still pause
Skip all approvals	Runs through everything, including those two. The app’s own label is blunt: “Lucius never pauses, even for unsafe actions.” Use only for an Automation you trust and have tested

Inside a run, the Automation’s own mode decides — never your global permission table, in either direction:

• A tool you’ve globally set to always_allow does not skip review on a manual Automation.
• A tool still needs_approval everywhere else does not get asked about on an auto or skip Automation.

The one thing no Automation mode overrides is a tool set to blocked globally — it’s refused before the Automation’s own settings are even consulted, in every mode including Skip. See Permissions for the levels and Confirm cards for what a pause looks like.

The extra safety net, and where it stops

Two operations are treated as riskier than an ordinary write, because the model chooses which rows are affected and there’s no way to check the result afterward if it chose wrong:

• Deleting a filtered set of transactions
• Re-scanning your whole ledger for a payee rule

On Automatically approve, these two still pause for your review even though everything else sails through. On Skip all approvals, they don’t — that mode means what it says, with no exception carved out for these two either.

“Always allow” inside a run

Tapping Always allow during a run records that permission on that Automation specifically, not globally. Review or revoke it from the Automation’s own settings.

Cost

A run costs the same as an equivalent chat conversation, on whichever model that Automation uses, and shows up in your cost history tagged as that Automation — see Models and cost control.

Before you turn one loose

An unattended run has the exact same reach as a conversation you’re watching. What actually protects you is upstream of the Automation’s own settings: the permission level of every tool it might call, plus the approval mode you choose. Read both before pointing an auto or skip Automation at anything that writes money. Scheduled runs also have known reliability gaps today — see A scheduled Automation didn’t run.