Replit, Softr and Bubble in September 2026: MCP Everywhere and Agents With Database Access
Disclosure: links to Replit, Softr and Bubble on this page are referral links. If you sign up through them we may earn a commission, at no extra cost.

Disclosure: links to Replit, Softr and Bubble on this page are referral links. If you sign up through them we may earn a commission, at no extra cost to you. Every item below comes from each platform’s own changelog or update post, read on 26 September 2026.
Read the September releases from Replit, Softr and Bubble side by side and the same two moves show up on all three. First, each one is opening itself to outside AI assistants over MCP, so you can build in one tool from a conversation in another. Second, each is handing its own agent more access to live project data. Taken together that is a genuine shift in how these platforms work — and a question about what you let an agent touch.
The headline from each
Replit: four changelog drops in a month, ending 25 September with apps for Meta devices, inline charts in the conversation, and GPT-6 Sol, GPT-6 Luna Fast and Claude Opus 5.5 as model choices.
Softr: a three-day MCP rollout in early September — build apps, custom blocks and workflows from Claude or ChatGPT, on every plan, with no extra credits.
Bubble: its AI agent can now read and write the development database with your permission, and new responsive pages default to responsive layout at last.
Replit: the busiest month of the three

| Date | What shipped | Who it is for |
|---|---|---|
| 25 Sep | Apps for Meta devices with device-specific agent instructions and QR previews; app creation through Muse; inline charts in the conversation; Airwallex sandbox over MCP; GPT-6 Sol, GPT-6 Luna Fast and Claude Opus 5.5 as model options | Everyone building with Agent |
| 25 Sep | Access requests when you hit a usage limit or restriction, off by default and enabled per workspace; design systems and slide templates now private by default under workspace permissions | Enterprise admins |
| 18 Sep | Custom API connectors in beta — point Agent at an API outside Replit’s integration library; audit logs expanded by more than 65 events | Pro and Enterprise |
| 11 Sep | MCP client integration, so a desktop AI client can create, inspect, update and publish Replit apps; Databricks Apps with Lakebase managed PostgreSQL; Compliance API can retrieve prompt text | Developers and compliance teams |
| 4 Sep | Scheduled database backups with one full restore point a day for production; Project Analytics covering visitors, pages, sources, countries and devices | Anyone running a live app |
| 28 Aug | Intelligent model routing — Auto mode picks the model balancing quality and cost; Growth Skills for Apollo, Clay, Stripe and PostHog | Cost-conscious builders |
The two most practical items are the least flashy. Daily database backups close a gap that used to make people nervous about running anything real on Replit, and Project Analytics means you no longer bolt on a third-party script to answer “is anyone using this?”. The model routing matters too: Auto mode is the difference between paying Opus prices for a one-line change and not.
Full details are in Replit’s changelog, and our Replit review covers the core product. You can try Replit through our link.
Softr: MCP on every plan, and agents inside your apps
Softr shipped its MCP support across three consecutive days in early September, and the notable part is the licensing: it works on all plans, and Softr states it needs no additional AI credits. That is unusual — most platforms put AI access behind the upper tiers.
- 6 September: build apps from Claude, ChatGPT or another MCP client.
- 7 September: Vibe Code Block MCP, for generating custom blocks mapped to your data sources.
- 8 September: Workflow MCP, including migrating workflows from other platforms by describing or screenshotting them.
Before that, on 26 August, came Softr Agents in two forms: App Agents that sit inside your app and answer user questions or perform tasks while respecting that user’s permissions, and Scheduled Agents that run recurring background jobs. Both are on all plans. The same day brought two-factor authentication for app users, optional or enforced, and Workspace Teams for restricting project visibility and inviting guests — that one is Business and Enterprise only.
Our Softr review has the wider picture; Softr has a free plan if you want to test the MCP server.
Bubble: the agent gets database access
Bubble’s official release notes page has not been updated since late July, so the substance now lives in its monthly community update. September’s list is long; these are the parts that change how you work:
- The AI agent can read and write your development database when you grant permission, with a setting to auto-approve actions or require approval each time. It can also install plugins, and uses a fine-tuned model for generating expressions.
- New responsive pages default to responsive layout instead of the old fixed layout — a small default with years of bad habits behind it.
- Mobile: a pre-formatted “Sign in with Apple” button that handles the OAuth flow, universal links, and workflows in generated mobile apps.
- Editor: a
:converted to dateoperator and a “formatted as” option for booleans; merges are around 30% faster in Bubble’s testing. - Enterprise: a hub for dedicated environments, and read replicas in private beta on dedicated plans.
Announced but not yet shipped: a mobile plugin editor beta, a web preview simulator, a workflow-engine overhaul with branching and looping (in alpha signup), and a Bubble MCP interface — which will put Bubble on the same footing as the other two here. Our Bubble review covers the platform itself, and Bubble is free to start.
The part worth pausing on
Between them, these releases mean an AI assistant running on your laptop can create and publish apps on your account, and a platform agent can write to your database. Both are genuinely useful. Both are also access decisions that are easy to make by clicking through a prompt.
Three habits worth adopting while the tooling is this new: keep agent write access pointed at development databases rather than production, leave approval-required on until you trust the pattern you are automating, and — if you are on a team plan — check who else in the workspace inherits the connection you just authorised. Replit’s expanded audit logs and Softr’s permission-respecting agents exist precisely because vendors know this is the risk area.
Which of the three moved most?
| Platform | September in one line | Best new reason to look again |
|---|---|---|
| Replit | Shipped weekly, across product, enterprise and models | Daily production database backups and built-in analytics |
| Softr | Made MCP free and universal across plans | Build and edit Softr apps from Claude or ChatGPT at no extra cost |
| Bubble | Gave its agent real database access and fixed a decade-old default | Responsive by default on new pages |
The verdict
If you already build on one of these, September is worth twenty minutes of your time: Replit for the backups and analytics, Softr for the MCP server, Bubble for the agent’s new reach. If you are choosing between them, none of this month’s releases changes the fundamentals — Replit for people who can code, Softr for internal tools over existing data, Bubble for complex logic without code — but all three are converging on the same idea, that the place you build should be reachable from whatever assistant you already talk to.
Sources: Replit’s changelog entries for 28 August and 4, 11, 18 and 25 September 2026; Softr’s what’s-new page and blog for 26 August and 6–8 September 2026; Bubble’s September 2026 monthly community update. All read on 26 September 2026.
Related guides
- Replit Review 2026: AI Chat, App Builder, Design Mode, Mobile Apps, and Deployment Features
- Softr Review 2026: New AI Agents, Secure Business Apps, Workflows, and No-Code Database Updates
- Bubble.io Review 2026: AI App Builder, Mobile Apps and Pricing
- YouWare Review 2026: Pricing, YouBase, and Whether It Is Still an App Builder
- Replit Agent benchmarks: 72% on DeepSWE at $2.11 a task
- MeDo now bills backends by size, daily
- Typeform’s MCP server: build, theme and publish forms from Claude or ChatGPT
- 21st.dev review: 12,000 shadcn components, an MCP server and no monthly plan
- Replit’s October 2026 update: GPT-6.1 Sol, Claude Sonnet 5.5 and the Jev decision model
