# Replit Agent Playbook: Modes, Plan Mode and Lower Costs

> How to work with Replit Agent: Free Mode vs Power and Max, Plan Mode, checkpoints, parallel tasks, testing and budgets that keep effort-based costs down.

_Source: https://professionalstoolkit.com/articles/replit-agent-playbook — The Professional's Toolkit · updated November 2026_

---

Replit's AI is **Replit Agent**: you describe an app in plain language, and Agent plans it, writes the code, tests it in a real browser and publishes it from the same cloud workspace. The current generation, Agent 4 (March 2026), builds web apps, mobile apps, slide decks, data visualizations, animations and 3D games inside one project. Free Mode routes each request to a model for you at no cost while your plan's allowance lasts, and the paid Power and Max modes spend your dollar credits by effort. Core costs $20 per month billed monthly, or $18 per month billed annually, and includes $20 of monthly credits. The habits that save the most: plan before building, stay in Free Mode for routine edits, and set a budget on day one.


**3 modes** — Free Mode, Power and Max on Starter, Core and Pro (Starter: Free Mode only); only the paid two spend credits, and Auto works inside them.



- **7 artifacts** — Per project, at most 1 mobile app; all publish together
- **1 / 10** — Parallel Agent tasks on Core / Pro (Starter: none)
- **7 or 28 days** — Production restore window: 7 days by default; Pro or Enterprise can choose 28


## What Agent 4 hands you at the end of a prompt

Agent's output is a **project**, and each thing it builds inside that project is an artifact. One project holds up to seven artifacts, at most one of them a mobile app, and they all share one database, one backend and one deployment. That has a practical consequence: when you publish, every artifact goes live together, and you cannot push one alone. An artifact also cannot be moved to another project later, so decide early whether a slide deck or dashboard belongs beside the app or in a project of its own.

Every web app Agent produces is full stack. It adds API routes and a PostgreSQL database when the idea needs them, and Replit handles hosting and SSL. How much of that you can reach depends on the build type. The free Starter plan gets **Lite build only**, a fast build experience for quick edits and targeted changes; Full build needs Core or Pro. Replit Design, the visual canvas, works on every plan, but a mockup there has no server and saves no data. Turning a design frame into a working app with a backend is a Core feature.


> 💡 **Write rules into replit.md:** Agent reads replit.md at the project root as standing context. Put your stack, naming rules and files it must not touch there before the first big build.


## Free, Power, Max and Auto: choosing who does the work

Replit organizes Agent into three modes, with Auto as a setting inside the paid two [1]. Free Mode exists on Starter, Core and Pro but not on Enterprise, which gets a single Auto mode instead. It is an allowance, not a plan. Core's card promises up to 30 hours of chat and up to 60 projects in Free Mode, Pro advertises even more, and on both paid plans the allowance resets every five hours under a weekly limit.

| Mode | Best use | What it spends | Model choice |
|---|---|---|---|
| Free Mode | Chatting, ideating, everyday building and edits | Nothing, within the plan's allowance | None: intelligent routing only |
| Power | Most paid work; Replit's recommended default | Monthly credits, by effort | Auto, or pick a model (Core and Pro) |
| Max | Complex, production-grade changes and long builds | Monthly credits, by effort | Auto, or pick a model (Core and Pro) |

Your Usage page shows how much Free Mode allowance is left and when it refreshes. When a request needs more muscle, Replit asks you to confirm the paid step for that request only, then drops back to Free Mode afterward. Decline, and nothing is charged. Starter accounts cannot escalate at all; Power and Max begin on Core, so a free account simply waits for its allowance to return.

Which model answers you? Replit's August 2026 launch post said Free Mode ran on OpenAI's GPT-5.6 Luna, yet the docs describe routing rather than one fixed model. In Power or Max you can keep Auto, which splits work across models and effort levels; Replit says that in its testing Auto matched the previous Max Mode's quality at 65% lower cost (August 2026). Or pick a model yourself, trusting the list inside your account, because model names change often.

Two finer controls sit beside the model. Each model carries an **Effort** dial from Low to Max, and higher effort can cost up to about twice as much on the hardest tasks. Compare models (beta) runs one prompt on up to four models, priced like four separate runs. Fast model variants require Pro.

## Plan before Agent touches a file

Plan Mode is the single best brake on wasted work, and it starts on Core. In it, Agent reads your project, searches documentation and answers questions, but it changes no file until you approve a plan [2]. You then choose to build in the current session or in the background. Replit creates a checkpoint when planning turns into building, and another as each task in the plan finishes, so every step can be reviewed and undone.

Planning is not free, though. Every interaction in Plan Mode is billed, text answers and task lists included. Replit's spend guidance still recommends it, because a plan uses fewer tokens than writing code, and a wrong plan costs far less than a wrong build. One limit catches people: Plan Mode lives in the Project Editor, not in Conversations. Context is the other lever. Agent writes a replit.md file at the project root and reads it as standing instructions, so put your stack choices, naming rules and *things never to touch* there. Paid plans add reusable workspace skills, and Pro adds workspace-wide custom instructions.


> 💡 **Open a fresh chat per feature:** Longer threads send more context with every message, so each request costs more. Start a new chat for unrelated work.


## Checkpoints undo code, not your customers' data

Agent saves checkpoints automatically at milestones. Each one captures files, conversation context, environment settings, Agent memory and database data and schema; it is also a Git commit and shows what that step cost [3]. Rolling back is one click from the Agent panel. If you then make new changes, Replit starts an alternate branch rather than overwriting the later history.

Here is the line that matters. A rollback leaves the database alone by default; you can tick an option to include the development database. **The production database is never restored by a rollback.** Replit introduced separate development and production databases in July 2025, every app now has both, and Agent only gets read access to production. Recovering live data needs a point-in-time restore, which keeps seven days by default, while Pro can choose up to 28 days. To take a published app back, roll the project to a checkpoint and publish again. Our [Replit database guide](/articles/replit-database) covers restores.


> ⚠ **Rollback skips production data:** Rolling back to a checkpoint never restores the production database; live data needs a point-in-time restore (7 days by default; Pro can choose up to 28 days).


## Running tasks side by side, and letting Agent test

Agent 4 splits bigger requests into tasks that run in isolated copies of the project, and nothing reaches the main version until you apply it. The plan sets how many run at once: none on Starter, one on Core, ten on Pro and up to 64 on Enterprise, with extras waiting in a queue. Replit said in March 2026 that about 90% of merge conflicts resolve without help. Parallel work suits independent pieces, such as a pricing page and an admin export, not two edits to one form. A canceled task is gone for good.

App Testing has Agent click through your app in a browser and fix what breaks, in a loop. It runs in Power or Max, is off in Free Mode, covers full-stack JavaScript and Streamlit Python apps only, and is billed like any other effort. An automatic Semgrep scan also checks the files Agent changes for risky patterns and hardcoded secrets, and paid plans add Security Agent for a deeper review. None of this certifies the code. Replit itself warns that Agent is probabilistic and may make mistakes.

## Keeping effort-based billing predictable

Effort-based pricing means a complex request costs more than a simple one, and Replit publishes no fixed price per prompt. Monthly credits are not a ceiling: once they run out, Power, Max, hosting and databases keep running and are billed pay-as-you-go unless you set a budget ([pricing guide](/articles/replit-pricing)). Replit's only examples are dated: simple changes typically under $0.25 (June 2025) and from $0.06 to several dollars per checkpoint (July 2025). Cost also grows with the project. Larger codebases and longer chat threads send more context with every message, so a fresh chat for an unrelated feature is cheaper than one endless thread.

You can watch the meter. Each checkpoint shows its cost, the chat header shows live costs while tasks run, and the usage dashboard can lag by up to 30 minutes. Simple Visual Editor changes to text, colors and spacing go straight into the code at no credit cost. Routines, which run Agent on an hourly, daily or weekly schedule in Power or Max, carry a per-run budget checked after each step, so a run can end slightly over it.

Set controls before Agent can loop. A usage alert warns you once spending passes an amount beyond your monthly credits. A **budget is a hard cap**: when it is met, Replit emails you and temporarily suspends every usage-based service, hosting and databases included [4]. Per-user spend limits are Enterprise only, and collaborators' Agent work spends the project owner's credits. Credits behave differently by plan. Core credits reset every month. Pro, which starts at $100 per month billed monthly, rolls unused credits over for two months. Credit packs on Core and Pro expire six months after purchase. Plan and pack details sit in our [Replit pricing guide](/articles/replit-pricing), and the free tier's limits in [is Replit free](/articles/is-replit-free).


## FAQ

**Can I use Replit Agent from my phone?**

Yes for prompting: Voice Mode works on all plans, including in the Replit mobile app, for recordings of up to 6 minutes. The Visual Editor works only in web browsers, and native mobile-app work (preview, build, App Store submission) runs in the Project Editor on replit.com, not in the Replit iOS app.

**Does Replit Agent work on code I already have?**

Yes. Agent works on any language or framework that runs on NixOS, and you can import from GitHub, GitLab, Bitbucket, Vercel, Bolt, Lovable, Base44, a ZIP file or Figma. Imports that need Agent to migrate the code can consume credits, and imports from Lovable, Base44 or Bolt bring code and schema but not data records, secrets or API keys.

**Can Agent be stopped from publishing or adding secrets on its own?**

Sensitive actions such as publishing, secret changes and outbound calls are approved by the builder, and in shared projects connecting an integration or adding a secret stays with the owner.

**Can I bring my own OpenAI or Anthropic key?**

Your own API keys, stored in Secrets, work for the AI features inside your app on any plan. Replit's built-in AI Integrations, which need no keys and bill credits, start on Core; on Pro and Enterprise an admin must turn them on first.


## The bottom line

Replit Agent rewards a deliberate workflow. Use Free Mode for chat and small edits, open Plan Mode before anything structural, accept a paid escalation only when the request truly needs it, and keep Power as your default paid mode, with Max for hard, multi-file work. Lean on checkpoints for code, but treat production data as separate, with its own restore window. Above all, set a budget and an alert on the first day, knowing the cap also pauses hosting. Starter shows how Agent thinks; Core is where Plan Mode, Full build and the paid modes begin.



## Sources

[1] Replit docs: Agent Modes — https://docs.replit.com/features/agent/agent-modes (2026-11)
[2] Replit docs: Plan vs. Build Mode — https://docs.replit.com/learn/plan-vs-build-mode (2026-11)
[3] Replit docs: Checkpoints and Rollbacks — https://docs.replit.com/features/version-control/checkpoints-and-rollbacks (2026-11)
[4] Replit docs: Managing Your Spend — https://docs.replit.com/billing/managing-spend (2026-11)
[5] Replit pricing page — https://replit.com/pricing
[6] Replit documentation — https://docs.replit.com/billing/plans/starter-plan
[7] Replit Terms of Service — https://replit.com/terms-of-service
[8] Replit Privacy Policy — https://replit.com/privacy-policy