Replit Agent: How It Builds Apps From Natural Language

Replit Agent is the AI builder inside Replit. You describe an app in plain language. Agent writes the code and sets up a database, then tests the result and publishes it. The current version is Agent 4, which Replit announced on March 11, 2026. Below: how a prompt becomes a published app, what each plan costs, and where Agent stores what it learns about your project. That last part decides whether you re-explain your stack every session.

What is Replit Agent, in one sentence?

Replit Agent is an AI builder in the Replit workspace that turns a plain-language description into a working app, then tests and publishes it without you writing code.

Replit's Agent docs frame it as an agent that acts on your project. Because it works on the project directly, it can run what it built and fix what broke without asking you. You can still open the code, but the default interface is a chat panel next to a live preview. The same agent also makes designs and slide decks from a prompt.

How does Replit Agent differ from a local coding agent?

People searching for Replit Agent often weigh it against running Claude Code or Codex on their own machine. Both take natural-language requests. The table shows what you set up and run yourself in each case.

Sources: the Replit Agent docs, the replit.md docs, Replit Memories, and Anthropic's Claude Code memory docs.

How does Replit Agent build an app from a prompt?

The Agent docs describe a short loop. You describe what you want in the Project Editor and can pick a project type, such as a web app or a mobile app. From there Agent handles code and infrastructure, and you refine in chat until it is ready to publish.

Replit's own building guide adds habits that make the loop work:

1. Be specific about what should change and what must stay the same.
2. Turn on Plan mode for anything bigger than a small fix. Agent writes an ordered task list and waits for your approval before touching files.
3. Add context. Paste screenshots or error messages, or annotate the Design Canvas.
4. Test in Preview the way a real user would. Then describe problems concretely.
5. Use checkpoints. The "Rollback here" button in the History panel restores your files along with Agent memory. You can choose to restore the database too.

Data gets the same treatment. You ask Agent to add a PostgreSQL database and describe what it should store. Per the database docs, Agent can edit the development database but cannot modify production. When you publish, schema changes from development are applied to production. Replit creates the production database at that point if your development database has data.

What changed with Agent 4?

Replit's Agent 4 post changes that loop in a few places.

It adds a design canvas. Agent generates design variants. You tweak the one you like and Agent applies it to the app.

Requests now go through a task system. Each one becomes a discrete task that you review before it merges. Sub-agents split big jobs into smaller pieces and resolve merge conflicts. Parallel tasks are limited to Pro and Enterprise users.

One project can also hold several output types. A web app and a mobile app can share one backend. Slide decks and data apps can live in the same project.

Agent can also query connected services from chat. The docs name BigQuery and Linear among them. The March 13 changelog adds that teammates can build on the same project at the same time.

Replit says Agent 4 helps teams ship production-ready software "10x faster." That is a vendor claim, and the post gives no benchmark behind it.

How much does Replit Agent cost?

Agent work is billed by effort. Each completed request produces one checkpoint, and small fixes usually cost less than large builds. Plan mode can bill for reasoning even when it changes no code. Replit asks for confirmation before a paid action starts.

The Agent docs list four modes. Free mode covers everyday work within your plan's allowance. For Core and Pro, that allowance resets every five hours and has weekly limits. Power and Max are paid modes with stronger models. Auto picks a model for you.

Plan prices on Replit's pricing page, accessed October 28, 2026:

The Starter row comes from Replit's plans page, which calls it no cost, and the billing docs. They also cap background tasks: one on Core and 10 on Pro. Pro credits roll over for two months.

How does Replit Agent remember your project?

Replit gives Agent three places to keep context between sessions.

The first is replit.md, a Markdown file in the project root. According to the docs, Agent creates it automatically for new projects and includes it as context on every request. Agent can update it as it learns about the project, and you can edit it too. Put your stack and coding rules there. It must sit in the root, because Agent does not detect copies in subfolders.

Replit also writes Memories while you work. The Memories docs describe user memory for your preferences in a workspace. Project memory holds the conventions and fixes for one project. You can also define custom memory. Memories are private by default. Replit does not use your memory in projects that collaborators can edit unless you turn sharing on. You can view and edit the memory file under Settings, then Customization, then Memory.

The third place is Skills: folders with a SKILL.md file in the project's /.agents/skills directory. The Skills docs say Agent sees each skill's name and description and loads the full file only when relevant. Skills follow the open Agent Skills specification, so the folder format works with other agents.

All three live in Replit. Memories stay in the workspace, and replit.md is a Replit convention that Claude Code, for one, does not load.

The bottom line

For a hosted app with a real database, Replit Agent needs the least setup of the options in the table above, since hosting and Postgres come with the workspace. Agent 4's Plan mode and task review make bigger changes safer. Start on Core and iterate in Free mode. Switch on Plan mode before large builds. Keep replit.md short and current. Check your usage settings in the first week, because complex requests can cost more than the plan price suggests.

How do you keep project memory when you build outside Replit too?

Many projects do not stay in one tool. A prototype starts in Replit Agent. Later the code moves to a laptop, where Claude Code or Codex edits it. Replit's Memories stay in your Replit workspace. replit.md travels with the files, but Claude Code reads CLAUDE.md or AGENTS.md. It will not pick up replit.md unless you point it there. Each local agent then builds its own notes from zero.

memU is a shared memory for those local hosts. It works with Claude Code and Codex among other agents. Every host reads from one memory backend. A scheduled background task slices new session logs into jobs. The agent decides what is worth keeping and writes it down as readable Markdown memory or as a reusable skill. memU itself makes no LLM calls: it stores and retrieves what the agent prepared. A standing instruction in CLAUDE.md or AGENTS.md tells the agent to retrieve before it answers. memU does not connect to Replit Agent itself, so it covers the local half of the project.

The hosted version is free and syncs across devices. To try it, get an API key from memu.so and send your coding agent one message:

Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU, API Key is YOUR_MEMU_API_KEY.

The self-hosted setup and the list of supported agents are in the memU GitHub repository.

FAQ

Can Replit Agent build a full-stack app from one prompt?

Yes. Agent writes the code and adds a PostgreSQL database when you ask. It tests the app before you publish. Larger apps usually take several rounds in chat, and Plan mode lets you approve the task list first.

Is Replit Agent free?

The Starter plan includes Agent with a daily cap. Core costs $20 a month and Pro $100 a month. Each includes an allowance for paid modes plus free usage that resets every five hours.

What is replit.md?

replit.md is a Markdown file in the project root that Agent reads as context on every request. Agent creates it for new projects and can update it. You can edit it to set coding rules and preferences.

Can Replit Agent change my production database?

No. Agent can edit the development database only. Schema changes are applied to production when you publish.

Does Replit Agent remember previous sessions?

Yes. It reads replit.md on every request and uses Replit Memories and installed skills. Memories are private by default and stay inside your Replit workspace.