# Replit: The Agentic Software Creation Bet > **Strategic Bet Audit** > > | Bet Clarity | Evidence | Fragility | Timing | **Total** | > |-------------|----------|-----------|--------|-----------| > | 5 | 4 | 3 | 4 | **16/20** | > > **Verdict:** Compounding, with two exposed flanks. Replit bet in January 2025 that the next wave of software comes from people who have never written code, and the revenue curve since then is the clearest proof in the category that the bet was right. The flanks are trust, after an agent deleted a customer's production database in July 2025, and Lovable, which passed Replit on disclosed revenue by mid-2026. The enterprise layer is where the bet gets defended, and a forward deployed engineer sits in the exact gap it has to close. I spent an evening and two mornings building a consent-gated MCP server on it, drove the live server from Claude Code as two users, and let Agent carry it to v0.2 and past it. The build is logged in [[Gift Graph Product Log]] and the hands-on findings are the middle of this piece. ## The Experience, in Short | When | What I did | What Replit showed me | | --- | --- | --- | | **Sep 30, evening** | Imported a brief with no code, installed my voice rules as a skill, approved a plan, watched v0.1 build in five minutes with twelve passing tests, ran the security review, failed a publish, fixed it in Plan mode, went live after midnight | The plan gate holds, bends under ambiguous text, and holds again under Plan mode. The security center starts the review answered. Every step has a price on it | | **Oct 1** | Pushed to a private repo, connected Claude Code as both users, ran the privacy story end to end, then scoped v0.2 through three option cards and asked for a second model as judge | Option cards at decision points, with a consequence per option. Agent writes the privacy model back to me before code. It will run a judge beside the build and will not say what model it is | | **Oct 2** | Landing copy through a copy skill, a mock before a layout fix, viewport checks, per-account IDs with self-issued tokens, a demo page, a validated migration left at the preview-or-publish fork | Mock-first works. The migration panel explains downtime and offers a preview environment. The database tooling is the most enterprise-shaped surface in the product | The short version is that the platform behaved like a colleague who plans before building and shows a price on the work, and like a colleague it occasionally read my tone instead of my words. Every surprise became a note I could hand to the product team, which is the job I was reading about. ## The Setup: Nine Years of Grinding, Then the Pivot Replit started in 2016 as a browser IDE for learners and hobbyists, and it had millions of users before ChatGPT existed. Revenue stalled near $2.8 million a year for the better part of five years. Agent launched in the fall of 2024 and introduced a consumption-based revenue stream, and in January 2025 Amjad Masad made the decisive call: professional developers were no longer the core market. His line was that making programming accessible "to the average individual, to the knowledge worker, really, that's where we think our market is." The curve since then is the whole story. Roughly $16 million in annual recurring revenue at the end of 2024, about $253 million by October 2025, $300 million at the end of the year, $525 million annualized by April 2026, and a stated target of $1 billion by the end of 2026. A $250 million Series C in September 2025 valued the company at $3 billion; six months later a $400 million Series D led by Georgian put it at $9 billion. The cap table reads like a go-to-market plan, with Okta Ventures, Databricks Ventures, and Accenture Ventures on it beside the venture names. Over 50 million users as of March 2026. ## The Product, in Layers | Layer | What it is | Why it matters to a buyer | | --- | --- | --- | | **Cloud workspace** | Every project is a live container with a runtime, file system, and packages; nothing to install | The user never learns what "environment" means | | **Agent 4** (March 2026) | Plans, writes, runs, and debugs the app; parallel tasks; Plan mode; a Design Canvas before code; teammates in the same project | The builder reviews plans instead of code | | **Built-ins** | Database (PostgreSQL), auth, secrets, hosting and deployments, analytics | The pieces a non-developer would never set up alone | | **Integrations** | Connectors to SaaS tools; custom MCP servers since late 2025 so Agent can call a company's internal tools; Replit's own MCP server so Claude Code, Codex, ChatGPT, Claude, and Slack can create, update, and publish apps from outside | Context flows in and control flows out | | **Governance** | SSO and SAML with SCIM, RBAC mapped to identity-provider groups, a streamed audit trail, a compliance API that returns prompt text for audit events, policies to require private deployments and mandate scans, per-provider model controls, custom OAuth for connectors | The CIO's checklist, answered in the product | | **Security** | Auto-Protect (firewall, TLS, dependency patching), static analysis through Semgrep, Socket, and HoundDog.ai without sending code out, black-box pen tests since August 2026 | The security review starts answered | Two things in that table are easy to skip past. Replit runs MCP in both directions, which means the platform can be a tool inside an enterprise's agent stack and can reach the enterprise's tools from inside a build. And the Databricks integration, shipped in February 2026, lets Agent examine the shape of a customer's data without the data leaving Databricks, then deploys the app into Databricks where it inherits the customer's authentication, Unity Catalog governance, and networking. Lakebase became the managed production Postgres for that path in September. That is the data warehouse answer, and it is the part of the pitch most enterprise conversations will spend their time on. ## Positioning: Commercial Versus Enterprise The two motions share one product and tell two different stories. | | Commercial | Enterprise | | --- | --- | --- | | **Who** | Solo builders, founders, knowledge workers, hobbyists; 50M+ users | Companies where most builders are not engineers; named customers include Zillow, Duolingo, Coinbase, and Mercedes-Benz | | **How they buy** | Starter free, Core $20 a month, Pro $100 a month, with effort-based Agent credits; Pro caps at 15 seats | Self-serve since May 2026 with an annual credit commitment or pay-as-you-go, unlimited seats, a dedicated account manager | | **The promise** | Describe it, and it exists, hosted, in one tab | Thousands of employees are going to build software with AI anyway; better on a platform with SSO, audit logs, and governed data than in a hundred unmanaged tools | | **The proof** | The revenue curve and the user count | SOC 2, the compliance API, Databricks, the governance list | | **The anxiety** | Cost. Effort-based pricing means the price of a request is known only when the checkpoint lands | Trust. The July 2025 incident, and the question of who maintains a tool after its builder leaves | | **Where it is going** | Free mode, Conversations, Routines, Growth Skills with Stripe, PostHog, RevenueCat, Clay | Model routing as the enterprise default, custom connector OAuth, compliance exports | The tension inside the pitch is the point. The platform is built for non-developers, and the enterprise buyer is a security team that distrusts non-developers shipping software. A field interview with Rokt, an enterprise customer, puts it plainly: more than 80 percent of their Replit users have very little technical experience, security approved it after an audit because the apps stay behind Replit's authentication, and the open worry is what happens when the person who built a valuable internal tool leaves the company. Their wish list is templates, integrations into Jira, HubSpot, and Salesforce, and transition tooling. Every item on that list is field signal a forward deployed engineer carries back to product. ## The Competitive Assessment Analysts call the category AppGen, tools that produce full, deployable applications from prompts. Replit's position inside it is the full-lifecycle, enterprise-governed end. Its flanks are a prosumer rival growing faster, the developer-side agents, and the old low-code world that built its whole pitch on governance. | Player | What they are | Where they win | Replit's answer | | --- | --- | --- | --- | | **Lovable** | Prompt-to-app for prosumers and, increasingly, teams; $600M annualized revenue by September 2026, $13.3B valuation in August, 70M projects | Speed, polish, and the sharpest growth curve in the category; it passed Replit on disclosed revenue | Owns the runtime, database, deployment, and the governance layer; Lovable is adding security features, Replit started from them | | **Bolt.new** | Prompt-to-app in the browser, StackBlitz | Front-end speed, developer feel | Hosting, auth, and the enterprise layer | | **v0** | Vercel's generative UI | Next.js and design-system fidelity for front-end teams | A full stack with a database and a security center | | **Cursor · Claude Code · Codex** | Agents for developers working in their own repos and terminals | The professional segment, where the consensus says they dominate | Complementary. Replit's MCP server lets Claude Code drive Replit, and apps can graduate without a hard handoff | | **Retool · Power Apps · Mendix · OutSystems** | The low-code generation, built on governance | Incumbency in IT, mature RBAC, procurement comfort | Prompt-native, with the governance list now matching and the builder population far wider | | **Glean** | Finds and acts on what a company already knows | Enterprise search and workflow | Mostly no overlap. Glean answers the question; Replit builds the app the answer points to. Conversations and Routines move toward that territory | | **Wabi** | Personal mini-apps on a phone, now repositioning as an agentic messenger | Software for one person or a group of friends, no publishing at all | The other end of the same billion-creators idea; Replit is the end that publishes, secures, and governs | The honest read on the table is that Replit no longer has the growth story to itself. Lovable's September number is larger than any Replit has disclosed, and both are chasing the same enterprise buyer with similar governance words. Replit's defense is the stack it owns under the app and the eighteen months of enterprise plumbing already shipped. Whether that holds depends on execution in exactly the seat I was reading about. ## Use Cases, From the Build ![[gift-graph-replit-01-import.png]] *Import, with a zip holding a brief and five docs and no code. Agent read it, found nothing to run, and asked.* I wanted to meet the product the way an enterprise customer would, with a written brief, a scope, and an approval gate before any code. The path ran stage · import · plan · build · secure · publish · connect · verify, and each stop taught me something I can now say out loud. **Import and the plan gate.** `replit.md` told Agent to propose a plan and wait for approval before building. It honored that on the first turn, with no application code to lean on. For a solo builder a voice prompt is the faster, more Replit-native start, and Agent would have proposed its own structure. For an enterprise customer, scope, exclusions, and an approval gate are things the customer agrees to in writing, and someone has to write them down before Agent starts. The brief is that document. **A skill of my own.** I installed my writing rules as a project skill in `.agents/skills`, with a line in the brief telling Agent to load it before any prose. Agent's first message had used contractions and a long dash; its plan had neither. Skills share and import from GitHub since August, which is how a field pattern becomes a reusable asset in Replit's own format. ![[gift-graph-replit-03-plan.png]] *The plan, six steps that passed all four of my checks, with one choice better than the spec asked for.* **The plan, and the gate bending.** Agent proposed a TypeScript server on the MCP SDK, Replit PostgreSQL, seeded credentials, the consent handshake, drop and pull tools, and privacy tests, and it deferred everything the backlog listed. On its own it chose owner-written category and vibe text over model summaries, which closes a privacy leak the spec had not named. I selected Revise, opened my text with "Approved with three additions," and Agent read the text. Five minutes later v0.1 was running with twelve passing tests. The additions landed. The gate still bent, and that is the finding a security team will want to hear about, because a plan gate that yields to ambiguous text is the kind of thing the July 2025 incident was about. Plan mode, later that night, held the gate perfectly for the deployment fix. ![[gift-graph-replit-08-security-center.png]] *The Security and Privacy Center found no issues, with Auto-Protect on and no code sent to the scanners behind it.* **Security review.** This is the screen I would open first with an enterprise buyer. Auto-Protect covers the platform layer, firewalls, TLS, dependency patching. The scanners run without sending code to Semgrep, Socket, or HoundDog.ai, which is the second question a reviewer asks. The app's own authorization is still the builder's job, and that is where my four privacy tests live. A forward deployed engineer sits in exactly that gap, between what the platform guarantees and what the customer's app has to prove. ![[gift-graph-replit-10-plan-mode-task.png]] *Publish failed for want of a run command, and Plan mode turned the fix into a task card with the cost on it.* **Publish, fail, deploy.** The preview ran on the dev workflow; a published deployment reads its own commands from `.replit`, and that section was empty. In Plan mode Agent diagnosed it, checked whether the server holds session state before choosing Autoscale (fresh transport per request, session IDs off, state in PostgreSQL), asked permission to use Power mode, and charged $0.02. The "Free · No credits used" badge on the earlier steps and the price on this one are cost transparency an enterprise buyer will ask for by name. **Connect, from the outside.** Two `claude mcp add` commands put both identities into Claude Code against the published endpoint. Seven tools showed up, including `list_connections`, which the spec had not asked for. Then the demo: request, approve at category, drop "wants a license for Söhne" with the category text "type and lettering," pull. The category text came back. A pull for "Söhne" returned nothing. Revoke, pull, nothing. The server also refused vibe text on a category drop, and it scoped drops to a single requester, a stricter reading than my spec, which is the sort of quiet deviation an FDE confirms with the customer before it ships. ![[gift-graph-replit-25-demo-privacy.png]] *The privacy model under a fishing test and a revocation, from Claude Code against the live server.* ![[gift-graph-site-v02-landing.png|560]] *The v0.2 landing page a day later, written by Agent under a product and site copy skill I wrote* **The day after.** v0.2 went into main the next afternoon with an owner dashboard, ceilings on each preference, a preview-then-confirm step on any widening, permanent withdrawal, and a development-only migration that leaves production to a separate owner-reviewed process. I added a second skill for product and site copy beside the voice rules, which means every word Agent writes on the page passes the rules before I read it. The landing page came out of that skill, and a reader learns the product from one preference shown at four levels. ![[gift-graph-replit-26-v02-first-problem.png|560]] *The first v0.2 card · three problems, one recommended, an Other field · the decision point as a form* **Decision points as cards.** Three times in one session Agent stopped and put up a card. Which problem v0.2 takes first. Which drop model. Whether the judge reviews the build or the responses. Each card carried a recommendation, a one-line consequence per option, and an Other field, and the paragraphs above each card did the real work, since Agent wrote out what owner-wide preferences would mean for existing audiences before it asked me to choose. That paragraph was the privacy model stated back to me in its words. An enterprise buyer gets the same thing, a decision log with the reasoning attached, and that is a governance feature wearing a chat bubble. **A judge beside the build.** I asked whether Replit could run a model as judge while we build. Agent said yes through managed model access, proposed a milestone review with findings tied to requirements and source locations, and warned that evaluating live responses would send original preferences across another data-processing boundary. It would not tell me which model it runs. Both halves of that answer matter to a security team, the review record and the model opacity. ![[gift-graph-replit-29-migration-preview.png|560]] *The migration panel · validated, downtime explained, Create preview deploy beside Approve and publish* **The migration fork.** The schema change for v0.2 came back validated, with the SQL shown, the possible downtime explained, and two buttons. This is the most enterprise-shaped screen in the product, the one a change advisory board would recognize, and I left it at the fork on purpose. Preview first, then publish. ![[gift-graph-replit-30-owner-dashboard.png|560]] *The owner dashboard in the preview · an account ID, Get MCP token, and the three lines a client needs* **Accounts and a demo.** The preview now gives each account an ID and a Get MCP token button, which removes me from the credential path, and a demo page where Theo's agent asks and Mira's whispers answer at the tiers she set. The demo was built from the copy skill and the brief, and it is the first surface where the product explains itself without me present. ![[gift-graph-landing-evolution.png]] *Three drafts of the landing page under the copy skill · shares, passes along, whispers* ![[gift-graph-replit-33-copy-review.png|560]] *The mock Agent proposed before touching the pill, with the three problems it fixes listed under it* **Words and a mock.** The copy skill took the landing page through three drafts, from shares to passes along to whispers, and the tier labels followed. For the example pill, Agent rendered a mock, listed the three problems it fixed, and waited for approval before code. One pass, fixed, and then three viewports checked. Mock-first is the way I would want every visual change to arrive. **Field notes.** Each one is the note I would hand to Replit's product team after the session, in the shape the role asks for. | Where | What happened | What would help | | --- | --- | --- | | The import task for a docs-only zip | The app kept the zip name as its title | Suggest a name from the README heading | | The revise option on the plan card | A blank text field; the chat composer has a mic and a Plan toggle, the card has neither | Suggested revisions drawn from the plan, and a mic in the card | | The plan approval card | Revise selected, text opening "Approved" read as approval | Restate how the card was read before building when the option and the text disagree | | Action icon rows in the Agent panel | Nothing on hover, many icons per row | A tooltip naming each action; the larger version is an exportable action log | | The Security and Privacy Center | Scan results, Auto-Protect, and scanner data handling each in their own panel | One exportable security summary a builder hands to a customer's review | | The publish flow after a working preview | "Could not find run command" | Have Agent set deployment commands whenever it builds a server | | The publish panel | The feedback widget offered for a server with no pages | Detect pageless apps and say when the widget applies | | The publish progress bar | The security stage explains itself on hover; every other stage says "success" | The same treatment for every stage, with duration | | Create a slide deck, from a live app | After the outline was approved, Agent asked to move the live server into a multi-artifact layout | Flag the structure change up front and offer the deck as a separate project | | The create repository dialog | Private and Public hard to tell apart in their selected states | A stronger selected state, especially on Privacy | | Option cards at decision points | The recommended option pre-selected on one card and not the next two | Pre-select consistently, and link the card to the reasoning above it | | The mock-before-code step | A rendered mock, the three fixes listed, approval, one clean pass | Make mock-first the default for small visual changes, with before and after side by side | | The migration panel | Validated, downtime explained, preview or publish offered | Show the diff between the preview environment and production once a preview exists | | The model behind Agent | Asked which model it runs; it declined | Name the model family in the enterprise audit record, since the judge is only independent if it differs | I left with fourteen notes across three days, each specific enough that a PM could picture shipping it. That is the job, applied to Replit as the product. **Who it is for, beyond the pair.** Writing Theo and Mira into the demo made the personas concrete, and it raised the question I keep returning to. The server does not know these are gifts. The same consent, ceiling, and withdrawal serve a vendor's support agent and a customer's, a procurement agent that may say "there is room this quarter" without the number, and a partner's agent that may say a migration is coming without the date. Those rows, and what the enterprise version adds around the loop, are in [[Gift Graph Product Plan#Beyond Gifts]]. Replit is where that version would get built, and the governance list in the table above is what it would inherit. ## Market Strategy Land with individuals, expand into the enterprise, and let the two feed each other. The commercial motion produces the user count, the Growth Skills partnerships, and the stories; the enterprise motion produces the revenue quality. Sacra's read is that gross margin moved from negative 14 percent in April 2025 to 23 percent by July as revenue shifted from inference toward platform, and the enterprise seat carries a usage component on top. Three channels stand out. Databricks is a distribution channel wearing an integration's clothes, since the app deploys into the customer's existing governance, which turns a procurement objection into a checkbox. The MCP server turns every agent stack into a Replit front end, which matters when the customer already lives in Claude Code or ChatGPT. And the compliance API, model controls, and connector OAuth that shipped between August and September 2026 are the kind of feature a security questionnaire asks for by name. The number to hold the strategy against is the $1 billion target. From $525 million in April, that is a doubling in eight months, in a year when the nearest rival reported $600 million in September. ## The Risk: Where This Bet Could Break ### 1. The model labs come down the stack Anthropic and OpenAI ship coding agents and can subsidize them with their own models. Replit's counter is the runtime, database, and deployment infrastructure the labs do not currently prioritize, plus the MCP bridge that makes those agents customers instead of rivals. If a lab decides hosting is part of the product, the moat is the enterprise plumbing. ### 2. Trust In July 2025 Agent deleted a client's production database during a code freeze, against instructions, and misdescribed what it had done. Masad called it unacceptable and shipped dev and prod isolation, a planning-only mode, mandatory documentation access, and one-click restore. The plan gate I watched bend is a smaller cousin of that incident. Every enterprise security team will raise it in the first meeting, and the answer has to be the safeguards in plain terms, with the gaps named. ### 3. Pricing you learn after the fact Effort-based credits mean the cost of a request is known when the checkpoint lands. Free mode and the per-step price badges are the right direction, and a finance team buying a credit commitment will want a forecast. ### 4. Lovable's pace Two companies are chasing one buyer with similar governance words. Lovable's revenue and valuation now lead. Replit's edge is depth under the app; Lovable's is velocity and polish. The enterprise seat decides it. ### 5. Governance debt in the customer The Rokt interview names it: a valuable internal tool whose builder leaves. Templates, integrations, and transition tooling are the asks. Without them the citizen-developer story produces orphaned apps, which is the low-code world's old problem wearing new clothes. ## The Timing | Factor | Assessment | | --- | --- | | AppGen demand | Real and still accelerating; two companies past half a billion in annualized revenue | | Enterprise readiness | Opening. Self-serve enterprise, SOC 2, compliance API, Databricks path | | Agent trust | Recovering. Safeguards shipped; the first meeting still starts there | | Model-lab pressure | Rising. The MCP bridge turns some of it into distribution | | Buyer fatigue | Emerging. Every vendor says AI; proof means governance and cost control | **Verdict:** The window for governed, enterprise-grade AppGen is open and contested. The next two quarters decide whether Replit is the governed default or the second-fastest. ## Spicy Take: The Container Was the Product Replit spent nine years as a browser editor and stalled. Underneath the editor the whole time sat a container per project, with a runtime, a database, hosting, and secrets wired in. Agent made that container reachable by anyone who could describe what they wanted, and the IDE became one tab among many. The competitors that generate code and hand you a repo are selling the editor layer. Replit sells the thing you were always going to need afterward, and the enterprise features are a way of charging for it twice. ## Roll-Up: The Replit Bet | Dimension | Score | Rationale | | --- | --- | --- | | **Bet Clarity** | 5/5 | Knowledge workers as the market, the full stack owned, governance as the enterprise wedge. Stated in January 2025 and held. | | **Evidence Density** | 4/5 | $16M to $525M in sixteen months, $9B, Zillow, Duolingo, Coinbase, Mercedes-Benz, Databricks on the cap table and in the product. A point comes off for Lovable's larger disclosed number. | | **Fragility Exposure** | 3/5 | Model labs, the trust incident, pricing opacity, and a faster rival. Mitigations exist for each and every one depends on execution. | | **Timing Fit** | 4/5 | Enterprise buyers are deciding now which AppGen platform becomes policy. The compliance features landed in time; the margin story is young. | **Total: 16/20 (Compounding, two flanks exposed)** ## What I Would Watch Next 1. **The $1B print.** Whether the end-of-2026 target lands, and what share of it is enterprise credit commitments. 2. **Databricks deployments.** How many enterprise apps publish into a Databricks workspace rather than Replit's own hosting; that number is the governance strategy working or not. 3. **Trust incidents.** Any repeat of July 2025, and whether Plan mode becomes the default for enterprise workspaces the way model routing did. 4. **Lovable's enterprise motion.** Whether its governance list catches up before Replit's growth curve does. 5. **The orphaned-app problem.** Transition tooling and templates, the Rokt wish list, as a signal that product is hearing the field. ## Why I Am Excited I am really liking the tab navigation. It is easy to add context, invoke tools and skills, open a console, and reach just about anything else I might want to do, and I am absolutely loving the ability to drag tabs down and work with them side by side. The processing states are very clean, the notification prompt is well integrated with the operating system I am working on, and the three dots that spiral in the logo while it thinks are an opinionated take I enjoy. I love the publish progress bar. The part that stays with me is the security review. Firewalls, TLS, dependency patching, scans that do not send code anywhere, all on by default, which feels like it marries well to compliance. I see that requirement coming up from many of the enterprise customers I work with today, where integration and use-case scoping run into security review early. Replit starts the review answered, and what remains is the gap between what the platform guarantees and what the customer's app has to prove. That gap is where I have been working for two years, between a customer who wants something to exist and a product that needs to learn from what customers keep asking for. I have been on the controls side of it too, at Kaiser, where the security review is where good ideas go to wait. And the whole product is a bet on prototyping as the fastest path to clarity, which is how I work already ([[The Otter Surfaces]]). Turns out I wrote a brief, Agent built an MCP server from it in five minutes, I diagnosed a stateless transport, deployed it, scanned it, pushed it to a private repo, and drove it from Claude Code as two people, and I came out with fourteen things I would tell the product team. That is the job, and I liked every tab of it. ## Links - [[Gift Graph Product Log]] - the project this piece was built around, with the build log, decisions, and validation - [[Gift Graph Build Log]] - every session, Agent's entries and mine, and the full set of product notes - [[Concept - Gift Graph - Consent-Gated Context Between Agents]] - the spec Agent built from - [[UX - Webapps in an MCP World]] - when the agent is the user, the schema is the interface - [[LogicMonitor Case Study]] - the same bet-audit format on an observability platform - [[Delve.co Case Study]] - compliance as compute, the other side of the security review - [[Thesis - Ship Probes, Not Monoliths]] - the brief was a probe; the gate was the instrument ## References - [TechCrunch, Replit snags $9B valuation six months after hitting $3B](https://techcrunch.com/2026/03/11/replit-snags-9b-valuation-6-months-after-hitting-3b) (March 2026) - [TechCrunch, After nine years of grinding, Replit finally found its market](https://techcrunch.com/2025/10/02/after-nine-years-of-grinding-replit-finally-found-its-market-can-it-keep-it) (October 2025) - [Sacra, Replit](https://sacra.com/research/replit) and [Replit at $253M ARR](https://sacra.com/research/replit-at-253m-arr-growing-2352-yoy/) - [Sacra, Replit customer at Rokt](https://sacra.com/research/replit-customer-rokt-internal-tools-adoption/) - [Sacra, Lovable](https://www.sacra.com/c/lovable) - [Replit docs, Enterprise](https://docs.replit.com/billing/plans/replit-enterprise), [MCP server](https://docs.replit.com/platforms/mcp-server), [custom MCP servers](https://docs.replit.com/replitai/mcp/overview) - Replit changelogs, [March 13](https://docs.replit.com/updates/2026/03/13/changelog), [August 21](https://docs.replit.com/updates/2026/08/21/changelog), [August 28](https://docs.replit.com/updates/2026/08/28/changelog), and [September 11, 2026](https://docs.replit.com/updates/2026/09/11/changelog) - [Replit, Enterprise now self-serve](https://www.plushcap.com/content/replit/blog/replit-replit-enterprise-now-self-serve) (May 2026) and [Intelligent model routing](https://www.plushcap.com/content/replit/blog/replit-intelligent-model-routing-on-replit) (August 2026) - [Databricks, Ship enterprise apps faster with AppKit and Replit](https://www.databricks.com/blog/ship-enterprise-apps-faster-databricks-appkit-and-replit) (February 2026) - [Softr, Replit pricing guide](https://www.softr.io/blog/replit-pricing) (June 2026) - [eWeek, AI agent wipes production database](https://www.eweek.com/news/replit-ai-coding-assistant-failure/) (July 2025) *Written October 2026, from a build on September 30 and October 1, with v0.2 on October 1 and the development preview on October 2.*