I tried to replace Convex. Eight platforms later, no

Convex stays. Most of what my AI has shipped since spring 2026 runs on it, and one honest day spent shopping for a replacement (eight platforms, pricing pages, agent docs, July 2026) ended where it started. Nothing else puts the whole backend in one TypeScript folder that an AI can hold in its head.

Why does Convex feel so much easier with AI coding?

Because the entire backend is code in one folder. Schema, queries, mutations, scheduled jobs, file storage, vector indexes and auth wiring are TypeScript files sitting next to each other, so the model changes a feature the way it changes any module: edit two files, let the types flow, click nothing.

convex/
  schema.ts       # tables, indexes, vector indexes
  users.ts        # queries and mutations
  billing.ts      # actions calling Stripe
  crons.ts        # scheduled jobs
  http.ts         # webhook endpoints
  auth.config.ts  # auth providers

That listing is not a simplification. It is the actual surface an agent has to understand before it can ship a feature, and it fits in one screen. When I ask for something like rate-limited invite links, the diff is three files in that folder, the generated types break loudly if the model gets a shape wrong, and there is no dashboard step waiting to be forgotten.

The tooling matches. Convex’s docs ship an official Claude Code plugin, one /plugin install convex@claude-plugins-official, with tools that read the dev deployment’s data and logs, run functions, and hooks that keep generated types in sync while the agent works [1]. The Codex plugin does the same through MCP and adds convex-expert and convex-reviewer subagents [2]. My agent does not guess what the backend looks like. It asks the backend.

Vector search is the tell for how the whole platform thinks. It is a schema declaration, a vectorIndex with a dimensions field (1536 for OpenAI embeddings), queried from server code [3]. No second database, no sync pipeline, no second invoice.

People call this way of building vibe coding. Fine. The label matters less than the property underneath it: the fewer surfaces a backend has, the less there is for an agent, or for me, to forget.

Where Supabase lost me

Supabase is the obvious counter-argument, and I gave it real projects before settling. In several ways it is the more capable system: real Postgres, real SQL, an escape hatch to the entire relational world. My problem was never capability. It was that the backend lives in five places at once: SQL migrations, row-level security policies, Edge Functions, storage rules, generated types.

Supabase’s own docs are clear about the discipline this takes. The golden rule is “never change the remote database directly”: every schema change becomes a timestamped SQL migration file, tested locally, committed, then pushed, and dashboard edits break the sync [4]. That is sound engineering. It is also five surfaces where an AI can update four things and forget the fifth, and in my projects it regularly did. On Convex the fifth place does not exist. Data still needs backfilling when shapes change, but that work lives in the same folder as everything else, not in a parallel artifact chain.

Is there a better backend for AI-first shipping right now?

Not for the way I build. The shopping happened on July 29, 2026: eight rivals’ pricing pages and agent docs, read with genuine intent to switch. Two came close. Appwrite as the mature all-in-one, InsForge as the same idea rebuilt agent-first. Both give up the one thing I refuse to give up: the entire backend in one place, in code.

The price column is the quiet finding: five of the nine floors sit between $25 and $30 a month. The category has agreed on what this is worth. The disagreement is entirely about shape.

Appwrite gets $25 a month for the whole organization and turns it into the biggest bundle on the list: 200,000 monthly active users, 150 GB storage, 2 TB bandwidth, 3.5 million function executions, daily backups [5]. The agent story is real too: skills for nine SDKs that install into Claude Code, Codex and Cursor [6], plus MCP servers that manage databases, users and files directly [7]. Two things keep it at number three. The backend is still a set of platform resources (buckets, permissions, schedules) you configure rather than one codebase you read. And Appwrite’s own vector search guide wires an Appwrite Function to OpenAI for embeddings and an external Pinecone index for the search itself [8]. That second platform is exactly what I went shopping to avoid.

InstantDB feels the most like Convex of anything I read: schema and permissions as code, an MCP server through which agents create apps and push schema and permission changes, instruction files for Claude Code and Codex [9]. The free tier is honest, 1 GB with unlimited API requests and no auto-pausing, and Pro is $30 for 10 GB, ten teammates and daily backups [10]. But it is a sync engine with auth and storage, not a home for server logic: the docs cover presence, cursors, permissions and an admin HTTP API, and hosted functions, cron jobs and vector search are simply not in them [11]. I would still be running a second platform for the server half. Disqualified, with respect.

Cloudflare is the value play: $5 a month with 10 million requests and 30 million CPU milliseconds included [12], and the parts are individually excellent, including Vectorize, a real vector database wired into Workers [13]. But the parts are the product: Workers, D1, R2, Queues and Vectorize each bring their own config, consumer auth is yours to bring, and the EU story needs a careful read. D1 supports an EU jurisdiction for where data lives, while Cloudflare’s docs state plainly that Workers may still access that database from anywhere; the constraint controls the database, not the compute [14]. Great unit economics. Not one system.

Firebase has quietly kept up, and auth is free to 50,000 monthly active users before Google Cloud Identity Platform pricing takes over [15]. It still feels like what it is, a suite of Google Cloud products with separate consoles and separate meters. My one-folder test fails at the front door.

Nhost packages Postgres, GraphQL, auth and storage from $25 with $15 of compute credits and unlimited users [16]. The workflow is the tell: in Nhost’s own local-development guide, one schema change produces a SQL migration file plus Hasura metadata in YAML, both committed and deployed together [17]. That is the Supabase ceremony with a different accent, and I already know how that story ends for me.

Encore.ts is the most interesting wrong answer. Declaring infrastructure in TypeScript and having it provisioned into your own AWS or GCP account is genuinely elegant, and the AI angle is thought through. But production starts at $49 per member plus $99 per cloud environment, before the $2.50 per resource and before the AWS bill itself [18]. That is $148 a month in platform fees before the first user shows up. It is a serious framework for a team with a cloud account, not a way to ship a product this week.

InsForge is the one I’m watching

One platform on the list is built on exactly my thesis. InsForge relaunched as 2.0 on March 9, 2026, calling itself the backend for agentic development, with the slogan “agent experience is the new developer experience” and the claim that 99% of operations on the platform now come from agents rather than humans [19]. The box matches the pitch: Postgres with pgvector, auth, storage, edge functions, realtime and an OpenAI-compatible model gateway across providers [19], cron-triggered functions on pg_cron [20], the whole thing Apache 2.0 and self-hostable with Docker [21]. Pro costs $25 with 100,000 monthly active users, an 8 GB database and $10 of compute credit, with small always-on instances from $5 [22].

So why is it number two and not number one? Partly because it is Postgres underneath, which means migrations and row-level security come back, agent-managed but present. Mostly because of age. Crunchbase lists a $1.5 million pre-seed from MindWorks Capital and Baidu Ventures [23], and MindWorks’ own portfolio page dates the company to 2025 [24]. Convex, for scale: a $24 million round led by a16z in November 2025 with Spark Capital co-leading [25], and by the company’s own April 2026 count, nearly 10,000 paying teams [26].

Convex, latest round

$24M

led by a16z, November 2025

paying teams on Convex

10,000

company figure, "nearly 10,000", April 2026

InsForge, pre-seed

$1.5M

MindWorks Capital and Baidu Ventures, 2025

The gap those three numbers describe is the whole decision. I do not park customer data on a pre-seed platform, however right its thesis. I do keep a tab open, and the Apache 2.0 self-hosting story is a real hedge for the day that changes.

What does Convex actually cost in the EU?

More than the pricing page suggests. Convex’s EU announcement puts the region in AWS’s eu-west-1 (Ireland) on every plan including free, but on paid plans EU resource pricing runs 30% above US rates, and the included usage that makes US plans generous, like Pro’s 25 million function calls [27], does not apply: EU deployments are billed on demand [28]. For a European product with real traffic, that is the number to model before choosing a region, and the least advertised line in Convex’s pricing story.

The rest of my complaints, for symmetry with the praise. Free and Starter deployments run the S16 class, 16 concurrent queries and 16 concurrent mutations, while Professional jumps to S256 [29], so load alone can force the $25-per-developer upgrade [27]. Convex Auth is still labeled beta, “it isn’t complete and may change in backward-incompatible ways”, and the docs steer serious auth toward Clerk, Auth0 or WorkOS, the last free to a million users [30]. And the one-folder elegance is also the lock-in: my backend is queries and mutations against a document model, not SQL that a reporting tool can read on a Tuesday.

Against all that, the entry price is hard to argue with: the $0 Starter tier runs real production apps for up to six developers with a million function calls included on US deployments [27]. Every one of the costs above is real, and I keep paying them, because the alternatives charge me during shipping. That is the hour I am least willing to sell.

What I’d actually do

Stay on Convex for anything that has to exist by Friday. Build one small test project on Appwrite the day mature auth at six-figure user counts becomes the actual bottleneck, because a flat $25 for a whole organization is a better deal than my per-developer bill the moment a team grows [5]. Check on InsForge in mid-2027: a real funding round and named production customers would change my answer, and its self-hosting license means trying it is cheap.

The test that beat every spec sheet, if you want to run it on your own stack: open the repo after a week away. If reading one folder tells you where everything lives, the platform passes. The last piece I wrote was about giving AI models the right amount of authority. This one is about giving them the right building. Small surface, everything in reach: it turns out that is the spec for both.