Actions

The chat header has an Actions button. It opens a list of actions for the agent, each on one row with its icon and name, and an “i” button that explains what the action does and why you would use it, or why it isn’t available now.

The list shows the built-in actions that are switched on, then your own actions, sorted into groups when Settings has groups.
While actions are running, the Actions button shows how many in a badge, and each running action says so in the list, with how long it has run.
An action can’t be started again on the same agent while it is still running: its row is disabled and says so.
No action runs on a single tap: each asks for confirmation, or opens its own dialog, first.
History, in the list’s top line, lists the agent’s action runs, newest first, each with its state, start time, duration, result and a link to its card.

What a run shows

Every action puts a card in the agent’s chat. The card shows the action, when it started, and each step as it goes, with its time, counting live. At the end it shows the total time and the result, or why it failed. An action that sends the agent a request shows that request inside its card, folded under Request. A run whose agent left sub-agents working ends only once they are done too.

Built-in actions

Conversation

Ignore my last message: for a message, or an action’s request, sent by mistake. After you confirm, the agent is asked not to act on your latest message or action request in this chat, to stop and undo what it already started for it, and to cancel any task it recorded for it. The chat shows a card naming what to ignore. It waits for the agent’s turn like any message; it never interrupts it.
Learn language: opens a small dialog of check marks for the kinds of mistakes the agent points out in your messages: Spelling, Wrong word, Verb forms and tenses, Plurals, Articles, Small words and Natural wording. It is on while at least one is ticked. The agent then points them out at the top of its answers (what you wrote → the fix), and nothing when there are none; never spacing, punctuation, capitals or chat shorthand.

System

Analyze system pressure: asks the agent to find what slows the host and the project’s container down now and what to do about it. It measures with read-only commands, samples more than once, finds the processes behind it, and reports the cause, how severe it is, and what to do, most effective first. It changes nothing without asking.
Update packages: updates the software installed on the project’s machine, never the project’s own dependencies. The dialog first checks what is installed (“Checking what’s installed…”) and lists only what it found: the system’s package manager, with Security updates and Other updates (or All updates for one that can’t do security updates alone), then other package managers and global tools, each unticked at first. One that needs root the account doesn’t have is shown disabled. The agent updates only what you ticked, never upgrades the release or removes packages without asking, and reports what was updated, what failed, and whether a restart is needed, asking first.
Clean temporary files: frees disk space, reviewed before anything is deleted. Tick The project (its working folder, ticked at first) and/or The system (the whole machine). The agent looks, read-only, and lists what it found in a card: each item with its path, size, what it is and why it is safe to delete, grouped as Project and System, with the total; items it isn’t sure of are unticked. Nothing is deleted until you press Delete the ticked. The card then shows what was deleted and the space freed. Files tracked in Git or not yet committed, databases, secrets and ProjectStart’s own files are never listed.

Project

Review the work: has the agent’s work reviewed by review sub-agents that the agent starts and briefs itself. The dialog chooses What to review (by default the uncommitted changes and the last commit; or the branch’s commits since main; or your own words) and what to check, all ticked by default: logic and correctness, performance, simplicity, real fixes rather than workarounds, security, duplication and reuse, tests, the definition, and localization. The agent then posts the findings, fixes the real ones and asks you about any it disagrees with.
Prepare for public release: for a project whose repository is to be published. Choose the checks (all ticked by default): Private information, Security of the code, Security of third-party code, Licence compatibility, Code audit, Localization, Fresh-clone build, Public documentation and Repository hygiene. With Private information, choose who the published history’s commits name: Keep the authors, or One author for every commit (a name and email), which rewrites the whole history, so every clone must be made again. The agent fixes the clear findings and asks about the rest; it shows you every history change before making it, and never force-pushes, deletes branches or makes anything public without asking.
Deploy: deploys the project to its production environment. The dialog shows the last deploy made with it and an optional Notes field. The agent first prepares the deploy the way this project deploys, deploying nothing yet, and reports a card with the commit, the steps, anything that can’t be undone, and what’s new since the last deploy (New, Fixed, Changed), in plain words. Nothing is deployed until you press Deploy now and confirm. The card then shows whether it was deployed, rolled back or failed.

App

Run a new web service: for a project in a container. Type what should run in What should run? (or tap an example such as “Run VS Code in the browser”) and press Run. The agent sets it up and gives you its link (see Web services).
Suggest product names: fill in What is it? (required when the project’s definition is empty) and optional Conditions, then Run. The agent suggests 8 to 12 names, each with a reason, in a card with a check mark per name. Check the ticked names has it check those for clashes (products, companies, trademarks, package registries, domains), each getting a verdict: clear, minor clash or taken. More names asks for a new list without repeats. You can check and ask for more as often as you like.
Check email deliverability: enter a Domain (filled in with the project’s when it has one). The agent checks, with read-only lookups and without sending email, the domain’s mail servers, SPF, DKIM, DMARC, reverse DNS, blocklists, MTA-STS and DNSSEC, and reports each as passed, warned or failed, with the exact DNS record to fix it.

Agent

Compact conversation: replaces the conversation so far with a summary, so each turn reads less and replies come faster (see Compaction). Its card shows the size before and after.
Clone agent: makes a new agent, named in a dialog, that carries on from this one as a checkpoint: it knows all this agent knew, gets a copy of its chat and its settings, but not its schedules, open questions or pending prompts. The chat switches to the clone.
Transfer agent: moves the agent to a linked project (To project lists only those). Its chat, files shown in it, settings and memory go with it; Its schedules and Its open questions go only when ticked. The working folder’s files stay with this project. When the other project has an agent of the same name, Name there asks for another. The agent’s chat shows a Moved from card in its new project.
Replace agent: a fresh session for the same agent, in the same chat, when its conversation has grown too long. The agent first writes a hand-off note (open tasks, work in progress, decisions); the new session starts by reading it.
Stop agent: ends the agent’s process and frees its memory. Its chat and session stay, and your next message starts it again where it left off.

Your own actions

Settings › Actions lists every action in the same groups as the menu.

Each built-in action has a switch to offer it or not.
New action adds your own: a name, an optional one-line description, its prompt and a switch. Running it opens a dialog showing the prompt, editable for this run; Run sends it to the agent as your request.
New group adds a group. Drag an action by its handle into a group or out of every group, and drag groups to reorder them; each drop is saved at once. Deleting a group leaves its actions in no group.
The built-in actions start in five groups: Conversation, System, Project, App and Agent, which you can rename, reorder or delete.
A project’s Settings › Actions has a switch per action for that project: an action switched off there isn’t offered to the project’s agents.
Every change is a History entry.

Review

Projects show their status in the sidebar; a status badge opens the latest attempt of its operation, with a live console for each step and, after a failure, what you can do next. Settings › History is the append-only record of what was done and by whom. It can be filtered (by project, among others), and Export CSV exports every matching entry, not just those shown.