The automatic CVE patching is the feature I did not expect. Most AI coding platforms in 2026 are competing on how fast they generate features. Replit Agent 4 added something different: a system that watches your projects for newly published security vulnerabilities and prepares a patch before you have to notice the CVE yourself.
That is a different category of capability. It is not making you build faster. It is reducing the chance your shipped app gets compromised by a known vulnerability while you were focused on something else.
Replit has had a reputation as a learning and prototyping platform. Agent 4 is a deliberate push toward production, and the security features are the clearest signal of that.
Agent 4: Parallel Tasks, Design Canvas and Plan Mode
Replit’s changelog introduced Agent 4 on March 13, 2026, pitching it as an “AI creative partner” that turns ideas into apps, designs, slides and more from plain language. The headline additions:
- Parallel tasks. Agent works on several tasks at once instead of one long sequential run.
- Design Canvas. You explore mockups on a canvas before any code is written.
- Plan mode. You map out a build before Agent starts generating.
- Team collaboration. Several people can build in the same project at the same time.
- Connectors. Agent pulls data into chat from BigQuery, Linear, Slack, Notion and others.
The parallel work gives faster initial generation on bigger apps and a clearer picture of what is being built before it is finished. You can see one task shaping the data layer while another works on the interface, rather than waiting for each step in turn.
Which model does the work depends on the mode you pick. According to Replit’s Agent modes docs, Free Mode uses intelligent model routing to select the model, Power Mode is “a smart balance of speed and capability”, and Max Mode is for complex work and larger codebases. Power and Max draw on monthly credits. The older Economy Mode is gone, replaced by Power Mode at the same price. If you care which foundation model writes your code, that routing is worth knowing about; the agentic coding guide covers why model choice matters less than workflow for most app builds.
iOS App Generation From Plain English
Replit launched natural-language mobile app building in January 2026. Describe the app, and Agent generates it.
The mobile docs are specific about how. Agent builds native apps with React Native and Expo, so the UI is platform-native components, not a webview, and one codebase targets iOS and Android. You do not need a Mac or Xcode, because Replit and Expo handle builds in the cloud. You publish from the Project Editor to TestFlight, then promote the build to the App Store in App Store Connect, and Apple reviews it before anyone can download it.
There are real catches. Submitting to TestFlight and the App Store requires an Apple Developer Program membership. Android apps can be built and previewed, but “Publishing to Google Play is not yet supported by Replit.” EAS CLI and EAS Build are not supported on Replit, and Expo Go only runs modules included in the Expo SDK.
Preview got better in April. The April 24 changelog added an iOS Simulator and Android Emulator streamed into the Preview panel, with hot reload and no Xcode, Android Studio or physical device. It is available on Core, Pro and Enterprise in Chrome, Safari and Chromium-based browsers, not Firefox. Simulators do not cover the camera, haptics, push notifications or GPS, so you still test those on a real phone with Expo Go.
The limitation is the same one that applies to all natural language app generation. Complex custom native features, deep OS integrations and performance-sensitive work still need a developer. But for apps that are mostly data display, forms and API calls, which describes most business apps, natural language generation covers the functional requirements.
Security Agent: The Feature That Earns Production Trust
The Security Agent is Replit’s most differentiated addition, and it shipped in the same April 24 update.
Replit calls it “a new AI-powered reviewer that audits not just your dependencies, but your entire project.” When you run a scan, it builds a threat model, maps routes and data flows, runs static analysis with Semgrep and HoundDog.ai, and then uses an LLM to decide which findings are actually exploitable in your app. Accepted findings are grouped into parallel tasks that Replit Agent can fix in one pass. It is available to all paid builders from the Security pane via Run Scan with Agent.
This is a meaningful improvement over running a static analysis tool and receiving hundreds of findings that need manual triage. The LLM-based filtering does the triage step and surfaces what matters.
CVE Auto-Protect takes this further, but it is opt-in, which the original version of this review got wrong. When it is turned on and a new critical CVE matches a dependency, Replit checks every project you own, has Agent prepare a patch in a background task, runs your tests and sends you a link to review it. The changelog is explicit: “nothing is published until you approve the fix.” I prefer that design. Silent auto-patching of production dependencies would worry me more than it would reassure me.
For context on why this matters: AI-generated code security vulnerabilities are a growing problem, with Veracode finding OWASP Top-10 flaws in 45% of AI-generated code samples. A platform that watches for known CVEs in generated code addresses one layer of that risk, and the vibe coding security risks breakdown covers the layers it does not.
App Monitoring and Production Diagnosis
Replit introduced App Monitoring for published apps in late April 2026, and it has since grown into what the docs call the Monitoring Agent.
Automatic checks watch three signals: application error logs, HTTP errors and request latency. You can simply ask Agent about production logs and metrics, or turn on Auto-fix, where Agent investigates alerts in the background using your metrics, logs and code and attempts a validated fix. An experimental Auto-republish setting goes one step further and publishes those fixes automatically.
I would leave Auto-republish off. The docs themselves warn that some changes, including migrations and index operations, may be risky, and “Auto-fix alone doesn’t authorize publishing changes to your live app” for good reason. Runs can also add Agent usage charges.
This is not a replacement for proper observability tooling on a serious production system. But for MVPs, internal tools and small SaaS products, having the builder also handle first-pass production diagnosis keeps everything in one place.
Connectors and Integrations
Agent 4 shipped with connectors for BigQuery, Linear, Slack and Notion, and the list keeps growing. The April update added a Plaid connector covering more than 12,000 financial institutions, where team admins set up credentials and each user authenticates their own bank account through Plaid Link with per-user access tokens.
Older projects are not left behind either. Projects created before Agent 4 can be upgraded to hold multiple artifacts, such as a mobile app or pitch deck, while keeping their URL, data, secrets and deployment. Replit said more than 200 projects migrated in the first 48 hours.
Replit Pricing in 2026
Replit’s pricing page lists three paid tiers:
| Plan | Monthly | Annual (per month) | Included |
|---|---|---|---|
| Core | $20 | $18 | $20 toward more powerful models, unlimited workspaces |
| Pro | $100 | $90 | $100 credits, up to 15 collaborators, database rollback up to 28 days |
| Enterprise | Custom | Custom | Custom terms |
(Source: replit.com/pricing, checked November 2, 2026.)
Free Mode is available on Starter, Core and Pro, and Core and Pro include a Free Mode allowance that refreshes every five hours. Power and Max Mode use your monthly credits.
Up to 15 collaborators on Pro is the line that matters for small teams. If your team is five people building internal tools, the comparison is one $100 plan against five per-seat subscriptions elsewhere.
Where Replit Agent 4 Has Limits
Replit generates complete applications from natural language, which means the output is optimised for the general case. Highly specific technical requirements, such as custom database configurations, non-standard auth flows or domain-specific algorithms, need manual extension or replacement of what Agent generated.
The iOS generation produces apps at the prototype-to-early-production level. Apps with deep native features, such as advanced camera handling, background processing or complex animations, need a developer to take the output further, and Android store publishing is still missing.
Replit is also a managed platform. If your team needs to run on-premise, in a specific cloud region or with specific infrastructure controls, that is a constraint.
Verdict
Replit Agent 4 is not the learning platform it was two years ago. Mobile app generation, the Security Agent, CVE Auto-Protect and the Monitoring Agent are production-minded additions that take the platform seriously as a place to build and deploy real software.
The Security Agent alone changes the calculus for teams using Replit beyond experimentation. Having a project-wide security auditor and opt-in CVE patching inside the platform you build on is meaningfully better than remembering to run external security tooling after the fact.
For teams looking for a single platform that handles web and mobile app generation, deployment and ongoing security maintenance, Replit Agent 4 is the most complete option in that specific category.
(Updated November 2, 2026: corrected several claims. Replit Agent’s model is now chosen per mode, with Free Mode routing automatically, not fixed to Claude 3.5 Sonnet, which Anthropic retired in October 2025. CVE Auto-Protect is opt-in and requires your approval before anything is published, not fully automatic. Removed an unverifiable “Pro Turbo Mode 2.5x” claim and other unverified pricing details, and replaced them with Replit’s current published Core and Pro prices. Removed unverifiable Razorpay MCP and Video Stack claims. Added sources.)