The Codelit agent

Write the outcome you want, for example Build me a habit tracker with a weekly streak chart, What changed on example.com/pricing?, or Plan a referral programme. Replies stream into the Thread. When the agent uses tools, its steps appear above the composer and you can stop it at any time.

• Slash commands such as /plan, /architecture and /app, @mentions, photos, and edits to the plan or architecture on screen keep their dedicated flows.
• Start a message with /agent to use the agent once when Send requests to the agent is off in Settings, Agent.
• The agent reads only websites you named in the conversation, or the exact pages a web search returned, so a page cannot send it elsewhere with your data.
• For current facts, prices and news, the agent can search the web. Each search costs a little from the request's budget, and the answer shows what the request cost.
• When you state a lasting preference or correct the agent, it may offer to remember it. Review and save opens Remember, where you can edit the wording first; nothing is saved without you.
• Rate an answer with Helpful or Not quite under it. Not quite asks what should be different next time and opens Remember with your note, so once you save it, later answers follow it.
• For work that needs a live session, such as clicking through a site or filling a form, or for desktop apps and files, the agent offers Browser or Computer. Choose Start it to open it with your request ready.
• In a private Thread, the agent uses the same saved goal, memories and skills as the Thread's chat.

Ask for a calculator, tracker, planner, dashboard, quiz or game and the agent builds a working app you can open from the link in the Thread or from My apps. To change one, ask the agent in a Thread and mention the app; your saved data is kept.

• Each app runs in a separate, sandboxed frame that blocks network requests, peer connections and form submissions, with no access to your Codelit account and no cookies.
• Data an app saves goes to your account, so it is there on your other devices. Each app can keep about 200 items.
• Delete an app and its data from the app's page. Deleting your Codelit account deletes them too.
• Add to team on the app's page shares it with one of your Codelit teams where you are an owner or admin. Everyone in the team opens it from My apps and uses one set of saved data: editors, admins and owners save, while viewers and reviewers use it without saving. Everyone in the team sees what is saved, so keep private details out. Only you, and the agent working for you, change the app's code. Stopping sharing, or someone leaving the team, ends their access at once. While the app is open, it tells you when a teammate saves, so you can reload before changing anything.
• Share a read-only copy with Share on the app's page. Anyone with the link can use the app, starting with empty data; your saved data stays private, and what visitors enter is not saved. The app's code, including anything written into it, is visible to anyone with the link. A link works only while you have Codelit Pro; Replace link retires the old one, and Stop sharing turns it off at once.
• A change updates the app in place. Versions on the app's page keeps the last five versions, each with the data it had then. Restoring one brings it back and keeps the current version in the list, so you can switch back. If the app is open in another tab when it changes, that tab asks you to reload before it saves again.
• If the agent changes an app after reading outside content, such as a web page or a connected app's data, it saves the change as a new app with a copy of the data, so the original stays as it was. The same happens in a Thread with answers from outside the agent, in a Thread started from a codelit.io link with a prompt (including Codelit for Mac's quick entry), for apps built from outside content, and for apps saved before in-place changes existed. To get an app that changes in place, ask the agent to build it again in a new Thread.
• Apps cannot show confirm boxes; a button that asks for confirmation does nothing and says so. Ask the agent for an in-page confirmation instead.

Anything that changes data in a connected app is shown to you as the exact action before it runs. Approve exactly this runs it once. Always allow approves it and lets the agent do that same action without asking next time. Decline changes nothing.

• Set Ask each time, Allow for 1 hour, Allow for 8 hours, Always allow or Never for each connected app action in Settings, Agent. A timed allow runs out by itself, even in the middle of a task, and then the action asks again: it ends at Ask each time, even if the action was set to Never before. Never removes the action from the agent entirely.
• Deletes, high-risk actions, and actions a service has not declared non-destructive always ask, whatever the rule.
• Always allow pauses once outside content is in the conversation: a web page, a connected app's data, an app's code, or the results of earlier agent steps in the Thread. Such text can try to steer the agent.
• After outside content, a request to start Product Plan, Architecture or Plan & Ship also waits for you to choose Start it. Builders always start from your own message.
• Actions run under a rule are recorded with that rule as their approval.

Ask for something on a schedule, such as Every weekday at 8:00, summarize new urgent Linear issues. The agent proposes a check-in with its time, time zone and instruction; Schedule it starts it. Each run reads and compares with its last report. You get an email only when something changed, and the full result is always in Check-ins.

• Check-ins cannot change data while you are away. A change one suggests is kept exactly as proposed in Check-ins, where you can review it and choose Approve exactly this or Dismiss.
• To watch a page, ask for example: check example.com/pricing every 4 hours and tell me what changed. Each run reads the page, and the agent only runs, costs and notifies when the text changed.
• You can watch a connected app the same way, for example: every 4 hours, tell me about new urgent Linear issues. The check reads the app and runs the agent only when the result changed.
• Run now in Settings, Agent runs a check-in once right away, so you can see what it does; its schedule stays the same.
• To react as things happen, open a check-in's webhook in Settings, Agent and create a webhook URL, then paste it into another service's webhook settings, such as GitHub, Linear, Stripe or Zapier. Events run the check-in at most once every 5 minutes and 12 times a day; events that arrive while it waits are read together in the next run, up to the latest five. Events are read as data, never as instructions.
• A webhook URL is shown once. Anyone who has it can start runs within those limits, so keep it private; Replace URL retires the old one at once, and Turn off stops it.
• To also require the sender's own signature, choose Require signature in the webhook panel with GitHub, Stripe or Linear and paste its signing secret. Events without a valid signature are then refused, and Stripe and Linear events must be recent.
• To post results to Slack as well, choose a channel with the # button next to a check-in. Posts carry the same short excerpt as email, without links; quiet runs and failures are not posted. If the check-in also runs on webhook events, whoever sends those events can influence what its reports say.
• Each run is limited to your per-run budget, up to $1. You can keep up to 10 check-ins.
• A check-in reads only websites you named in the conversation when you set it up. The approval card lists them, along with any page it watches.
• Pause, resume, turn off email, or delete a check-in in Settings, Agent. A check-in pauses itself after three runs in a row cannot finish, if your plan no longer includes the agent, or when you change AI access away from Within my budget.
• Each check-in keeps its latest ten results. Emails carry a short excerpt with links and addresses removed; the full result is in Check-ins.

Turn on Answer me in Slack in Settings, Agent, then message Codelit's Slack bot directly, or @mention it in a channel it is in. The agent answers in that Slack thread. In a direct message, reply in the thread to continue; in a channel, @mention Codelit again in the thread.

• Connect Slack in Settings, Integrations first. If you connected it before Answer me in Slack was available, disconnect it and connect it again, so Codelit knows which Slack account is yours.
• Codelit keeps your workspace's Slack bot token on its own server. It never reaches your browser, anyone else's, or a person's vault. Others in your workspace who connect Codelit use the same bot through their own Codelit accounts, but each person reads and posts only in channels they are in, and no one can read your direct messages with the bot through Codelit. In a Codelit team, Rooms and automations run on the owner's connected apps but use only Slack channels that the member using them is in too, so each member connects Slack themselves.
• In a channel, an answer is marked Only visible to you, because it draws on your own connected apps, and Slack removes it when Slack reloads. In a direct message, an answer is a normal message that stays in the conversation, once only Codelit can read it. Until then, for example while a bot token handed out by an older version of Codelit may still work, direct message answers are also only visible to you.
• Only messages from your own Slack account are answered. Messages from anyone else in the workspace, from bots, and edits are ignored.
• Each answer stays within the budget you choose, up to $1. Codelit answers one message at a time, up to 10 an hour and 30 a day.
• Answers in Slack work like check-ins: the agent reads, searches and builds, and a change to a connected app is kept as a suggestion. Review it in Check-ins and choose Approve exactly this or Dismiss.
• Slack shows you a short excerpt without links or addresses, because Slack opens links in a bot's message to preview them. Every full answer, with its sources, stays in Check-ins.
• The agent sees only your messages to it, not the rest of the Slack conversation.
• Turning Answer me in Slack off, or changing AI access away from Within my budget, stops answers there. Disconnecting Slack also turns it off; after you connect Slack again, turn it back on.