Replit Agent

Where Replit Agent sits

Replit Agent is the AI building partner inside Replit's browser IDE: you describe an application in ordinary language and it plans the work, writes code across the stack, provisions a database, tests what it built and hands you a deployed URL without leaving the tab. Agent 3, the generation shipping at time of writing, arrived in September 2025 and shifted the product from "impressive prototype machine" to something people finish MVPs with, mainly by extending autonomous run time and by having the agent test its own output in a real browser. For the buyer I care about — a founder, a student, an ops person automating internal tools — the appeal is not model quality in the abstract. It is that the entire delivery chain (code, database, auth, hosting, custom domain) lives in one subscription and one UI, so there is no deployment step to research and no stack of accounts to wire together.

What it actually does

The loop starts in a chat: you describe the app, Agent drafts an ordered plan (Plan mode exists specifically to review and refine that plan before any code changes), and then it builds. It writes frontend and backend code, sets up Replit's integrated PostgreSQL or SQLite, and works through discrete checkpoints — commits of shippable progress after which it summarises what changed and gives you a live preview. Checkpoints double as the undo mechanism: you can roll back to any previous one, which matters because agents build forward from assumptions. With App Testing on, Agent opens what it just built in a headless browser, clicks buttons, submits forms and checks API responses, then fixes what it finds; Replit claims this is 3x faster and 10x cheaper than computer-use style testing, which is a vendor claim, but the observable behaviour is real: fewer "submit button does nothing" handoffs. Max Autonomy lets a session run 200 minutes or more without asking permission at each step, and Agent 3 can also generate other automations ("Stacks") — a Slack bot, a Notion-connected workflow — from a description.

Deployment is part of the product rather than an afterthought: Autoscale for variable-traffic web apps and APIs (scales to zero, brief cold starts), Reserved VM for always-on work, and Static for pure frontends, with custom domains supported. Native mobile previews via Expo let you scan a QR code and test a build on a phone.

Billing is where buyers need to read carefully. Agent work is metered with effort-based pricing: paid modes (Power, Max, and Auto which picks between them) charge per checkpoint depending on task complexity, and there is a no-cost Free Mode with allowances that reset every five hours under weekly caps on Core and Pro. At time of writing, replit.com/pricing lists Starter at $0 (daily Agent credits up to a monthly cap, one published project, one background task), Core at $20 per month or $18 billed annually (with $20 of credit toward the most powerful models, unlimited workspaces, up to 30 hours of Free Mode chat and 60 projects), Pro at $100 per month or $90 billed annually ($100 of model credit, 10 parallel agents, up to 15 collaborators, 50 viewers, 28-day database rollback, premium support), and Enterprise custom (SSO/SCIM, audit logs, single-tenant options, static outbound IPs). Replit's own pricing FAQ confirms effort-based pricing is pay-as-you-go on top of the subscription.

My testing notes

I took a deliberately ordinary brief — a small client portal with auth, a Postgres data model and a dashboard — from prompt to deployed Autoscale URL in a single evening session. Two things stood out. First, the checkpoint cadence is the real editing surface: after each one I got a summary plus a working preview, and course-correcting at checkpoint three (a wrong assumption about how invoices relate to clients) cost a sentence, whereas the same correction after the tenth checkpoint would have meant unwinding several files of generated code. Second, the self-testing loop caught things I expected to catch myself: a dead form submit and an auth redirect that looped. It did not catch the missing rate limit on the public endpoint, and it did not catch an admin route that only checked authentication rather than authorisation — I had to ask for both explicitly. Third, cost: on a Core subscription with one focused project week, effort-based charges tracked the checkpoint count closely enough that the billing dashboard's running estimate was a fair predictor, but the default spend cap is off until you set one, and I can see how a long Max Autonomy session going the wrong direction produces a bad invoice. Set the cap on day one.

The collaboration story deserves a paragraph because it changed in February 2026 and a lot of published material is stale. Replit retired the Teams plan, folded its users into Pro, and pushed collaboration features down into Core: unlimited workspaces on paid plans, up to five invited people in a Core workspace, and on Pro up to 15 builders plus 50 read-only viewers, per-workspace usage tracking and database rollback to 28 days. For a small team shipping an internal tool, that is enough to run a shared project without paying per seat the way traditional IDE seats work; the cost risk sits in metered agent work instead. Replit also sells Agent modes as cost dials — Economy Mode is billed as roughly a third of the default agent per prompt, Power is the mid-tier default for real builds, and Turbo (Pro and Enterprise only) runs up to twice as fast at up to six times the cost. Having these as explicit modes rather than hidden model routing is the right design: it makes the bill a choice.

What it does not do

Agent is strongest on the Node/React/Postgres default path; ask for Rust, Elixir or a niche framework and quality drops visibly. Its context management is tuned for building forward — third-party guides put the practical ceiling around 15–20k lines before edits slow and errors multiply, at which point a local tool built for large codebases is the better home. The UI it writes is functional rather than art-directed: if brand-perfect design matters, plan a design pass afterwards. And treat everything it deploys as a first draft from a security standpoint; the agent will not volunteer hardening. The billing model also makes monthly cost genuinely harder to predict than a flat per-seat tool, and Replit's own footer concedes the agent's behaviour is probabilistic.

If you buy one habit from this review, make it this: set the spend cap in billing settings before the first long autonomous session, and treat each checkpoint summary as a review point rather than a notification to dismiss. The users who describe Replit Agent as expensive are, in the reports I read, mostly users who let Max Autonomy run unattended on an underspecified brief; the users who describe it as a bargain are the ones who corrected course at checkpoints and stayed in Free Mode for ordinary edits. The tool prices your supervision, which is unusual and, on balance, fair.

Verdict

Rated 4.1 overall — core functionality is deep for this category (autonomous multi-step builds, self-testing, integrated database and deployment), value is good at Core if you set a spend cap and poor if you do not, usability is the best thing about it (zero setup, browser-native), the docs and community are mature though human support is gated at Pro, and Replit's decade of running a cloud IDE is a real reliability signal. For founders and students who want an AI to build and ship a working app from a description, nothing in this price range removes more steps.

Sources

• https://replit.com/pricing
• https://docs.replit.com/features/agent/overview
• https://docs.replit.com/billing/plans
• https://replit.com/blog/pro-plan
• https://replit.com/blog/introducing-agent-3-our-most-autonomous-agent-yet
• https://codingagentwatch.com/tools/replit-agent/
• https://whatsthebestai.com/pricing/replit
• https://nisai.dev/guides/replit-agent-guide-2026
• https://dev.to/jovan_chan_9500711396d4e6/replit-agent-review-2026-from-idea-to-deployed-app-but-read-the-credit-bill-first-4cma
• https://aionx.co/ai-comparisons/replit-agent-review
• https://launchpad.io/blog/replit-pricing