To use Replit Agent, open a new chat on replit.com, choose what you want to build (a website, mobile app, slides or a dashboard), pick Free Mode, and describe one outcome with clear success criteria and limits. Agent plans the work, writes the code, starts the app and checks it. You test the result in Preview, ask for one focused change at a time, roll back to a checkpoint when a change goes wrong, and press Publish only when the preview works. The rest of this guide covers the choices that decide whether that loop stays cheap and under control.

We wrote this from Replit's own documentation, pricing page and blog as of November 11, 2026, not from a paid account, so treat menu names as current but subject to change. Agent 4 launched on March 11, 2026, and some Agent 3 features were renamed.

What Replit Agent is (and what it is not)

Replit Agent is a cloud app builder. You chat in a browser, and Agent sets up the project, writes code, provisions things like a database, runs the app and tests it, all on Replit's servers. Its documentation describes the target user as someone who does not need to code at all, and its output can be a web app, a mobile app, slides, a design, a data visualization or a mix of these sharing one backend.

That makes it a different tool from a coding agent that works inside your own repository on your own machine, such as Claude Code, or an AI editor such as Cursor. If you are not sure which camp you need, our explainer on what agentic coding is covers the plan, act and verify loop all of these tools share, and the best AI coding assistant guide compares the editor and terminal options. Replit's closest competitors are other prompt to app builders like Lovable.

Replit also has an early access workflow called Code with Replit that connects GitHub or GitLab repositories and prepares pull requests for your team to review. This guide follows the app builder. If you would rather write the code yourself in Replit's editor, see how to use Replit as an IDE.

Step 1: Choose a plan and an Agent mode

Two settings decide what Agent can do and what it costs: your plan and the mode you pick in the prompt box.

PlanPrice (Nov 11, 2026)What matters for Agent
StarterFreeAgent chat and Free Mode with a daily cap; no Plan Mode, no full build, no connectors; 1 published app that expires after 30 days
Core$20 a month, or $18 a month billed annuallyFree Mode allowance, $20 of monthly credits toward stronger models, Plan Mode, connectors, 1 active background task
Pro$100 a month, or $90 a month billed annuallyEverything in Core, $100 of credits that roll over for two months, 10 background tasks, database recovery up to 28 days

Sources: the Replit pricing page and the Replit AI billing docs.

The modes, per the Agent modes page:

  • Free Mode picks a model for you and costs nothing within your allowance. On Core and Pro the allowance refreshes every five hours and also has weekly limits. Starter has a daily cap.
  • Power Mode is the paid middle tier. It replaced the old Economy Mode at the same price.
  • Max Mode uses the strongest models for large codebases and hard problems.
  • Auto lets Replit choose the model for each paid task inside Power or Max.

Start in Free Mode. If a request needs more than Free Mode can handle, Agent may offer a paid "Allow once" option, and Replit asks you to confirm before any paid action starts. That confirmation is your main guard against surprise charges, so read it rather than clicking through.

Step 2: Plan before you build

For anything bigger than a single page, turn on Plan Mode first (the Plan toggle in the prompt box, or just ask Agent to make a plan). In Plan Mode Agent breaks the work into an ordered task list without touching code or data. The Plan Mode docs list four buttons on the finished plan:

  • Revise keeps refining the plan.
  • Cancel throws it away.
  • Build here approves it and builds in your current session.
  • Build in background approves it and runs it as a separate task.

Before your project's first checkpoint you will see a single Start building button instead of the two build options.

Two things catch people out. First, Plan Mode is not free on the paid path: Replit charges for planning answers and task lists even when no code changes. Second, background tasks run in an isolated copy of your project and wait for you to review and apply their changes, unless you switch on "Automatically apply changes". Agent 4 lets you keep planning in a new chat while the main build runs, and if two tasks touch the same files, Replit flags the conflict. Core allows one active background task and Pro allows ten, so on Core this is mostly a way to queue work, not run it in parallel. Our guide to single agent vs multi agent designs explains why parallel tasks need tight, self contained briefs.

Step 3: Write prompts with an outcome, criteria and limits

Replit's own first app tutorial teaches a five part loop that works well:

  1. Describe one useful outcome.
  2. Add success criteria you can check by looking at the app.
  3. Say what not to build yet.
  4. Test in Preview.
  5. Ask for one focused improvement.

A weak first prompt is "build me a booking app". A stronger one names the users, the three or four things they must be able to do, what done looks like, and what to leave out for now: no login, no payments, no database until the core screens work. Those limits keep the first build small, and small builds are cheaper because Replit prices paid work by effort.

When something breaks, describe the symptom instead of guessing at the cause. "When I submit the form, the success message shows but the old text stays in the fields" gives Agent far more to work with than "fix the form". Screenshots and annotated designs count as context too, and Replit's Build with Agent guide recommends saying what must stay the same so Agent does not rewrite parts that already work.

Step 4: Test in Preview, then use checkpoints

Agent tests its own work, but you should still use the app as a visitor would, including at phone width.

When a change makes things worse, use a checkpoint instead of arguing with Agent through ten more messages. Open History in the Agent panel, pick a checkpoint and select Rollback here. According to the checkpoints and rollbacks docs, a rollback restores your project files and Agent's conversation context to that point and removes everything made after it.

The database is the exception. By default a rollback does not change your database. To roll the development database back too, tick Database under the additional rollback options. A rollback never restores your production database; that needs a separate point in time restore, which covers 7 days on standard plans and 28 days on Pro. If your app writes real data, decide which of these you want before you click.

Step 5: Publish, and remember Preview is not your live app

The preview URL is private, temporary and meant for development. Sharing it does not give anyone access. To put the app online, select Publish in the top right of the project editor (or ask Agent to publish). The publishing guide walks through the options:

  • Domain: a generated .replit.app address you can rename, or a custom domain.
  • Access: Public, Password protected, Workspace only or Invite only.
  • Security: an optional deep security scan and a setting that blocks publishing when critical vulnerabilities are found. Turn both on for anything that takes user input.
  • Deployment type: Autoscale (adds machines as traffic grows), Reserved VM (always on), Static (files with no backend) or Scheduled (runs on a timetable).

Publishing makes a separate production deployment. Later edits in the editor do not touch the live app until you press Republish, which keeps the same URL. Test the public URL after every publish. Publishing works on every plan, but Starter's single app expires after 30 days, and machine size and traffic on paid plans draw on your credits. Our guide on whether Replit is free to deploy breaks down the cost of each deployment type.

Step 6: Give Agent standing context and tools

Three features keep you from repeating yourself, and Replit's docs on memories, custom instructions and skills explain when to use each:

  • Custom Instructions: rules you write that always apply, such as an approved library or a data handling requirement.
  • Skills: reusable procedures Agent loads when relevant. The idea matches the SKILL.md format in our agent skills guide.
  • Memories: context Replit saves from past work. Your explicit request and your Custom Instructions win over Memories, and you can review them under Settings.

To give Agent outside tools, connect an MCP server. Go to replit.com/integrations, choose MCP, and sign in to a listed server such as Stripe, Linear, Notion or Sentry. For anything else, the Connect via MCP docs describe adding a custom server by its HTTPS URL, with optional headers for an API key. Connected servers are available in all your projects, so only add ones you trust. If MCP is new to you, start with what MCP is.

How to keep Replit Agent costs under control

Paid Agent work uses effort based pricing: Agent makes at most one checkpoint per request, and the charge depends on how much computing the request used. Replit's effort based pricing post says simpler requests can cost as little as $0.06 while involved ones can run to several dollars, and that each new message costs more in a large project with a long thread. Some Agent features also call outside model APIs on your behalf, billed from your credits at the provider's public rate.

What helps in practice:

  • Do everyday work in Free Mode and check your remaining allowance under Settings, Usage.
  • Set usage alerts and a hard budget limit on the Billing settings page. Usage data can take up to 30 minutes to appear, so leave headroom.
  • Batch related small fixes into one clear request instead of five tiny ones.
  • Plan big features first, then build, so you do not pay for a wrong direction.

When Replit Agent is the wrong tool

Replit Agent is strongest when you want an idea running at a public URL without setting up a laptop, a database or hosting. It is a weaker fit when you already have a large codebase on GitHub with its own CI and hosting, when code must never leave your machines, or when you want to read and approve every diff. For those cases a local coding agent or AI editor fits better, and an AGENTS.md file gives that agent the same kind of standing instructions that Custom Instructions give Replit.

FAQ

Is Replit Agent free to use?

Partly. The Starter plan includes Agent chat and Free Mode with a daily cap, but not Plan Mode, full builds or connectors. Core and Pro add a Free Mode allowance that refreshes every five hours, plus monthly credits for paid Power and Max work.

Do I need to know how to code to use Replit Agent?

No. Replit designs Agent for people who describe what they want in plain language. Knowing some code still helps you judge whether a fix is right and spot when Agent changed something you did not ask for.

What is the difference between Plan Mode and Build mode in Replit Agent?

Plan Mode lets Agent brainstorm and produce a task list without changing code or data. With the Plan toggle off, Agent builds directly, which Replit calls Build mode. On the paid path both are billed.

Does a Replit checkpoint rollback undo database changes?

Not by default. A rollback restores files and Agent's conversation context. You must tick Database in the additional rollback options to restore the development database, and the production database needs a separate point in time restore.

How do I make my Replit app public?

Select Publish in the project editor, pick a domain and the Public access level, and publish. The development preview stays private, and later edits only go live when you select Republish.

Can Replit Agent work on my existing GitHub project?

Replit offers Code with Replit for GitHub and GitLab repositories in early access. It runs the app in a cloud sandbox and prepares pull requests for review, but availability differs by workspace.