Superblocks
Superblocks is AI-native. Your teams build in natural language, on an AI that learns your data, runs inside your own cloud, and stays under IT's governance.
An AI builder on top isn't a platform built for AI.
When AI is added on top of an existing platform, it behaves like a layer, not a foundation. The natural-language builder, app-building over MCP, and the newest models live in the cloud edition and in beta, and self-hosted deployments fall back to the older experience until features catch up. Superblocks was built AI-native from the core, so the same generation, governance, and code run in your own cloud today.
What makes Retool hard to love?
Five places the difference shows up when you build with AI.
- The AIAI sits in a separate builder, still in cloud beta.The AIThe development model across the whole platform.
- In your VPCNo AI-native builder when you self-host.In your VPCSame AI-native platform in your cloud, today.
- Your inferenceSelf-hosted falls behind on the newest models.Your inferenceYour Snowflake, Databricks, or Bedrock.
- Self-hosted footprintYou operate a multi-service stack and Postgres per VPC.Self-hosted footprintOne stateless Docker agent per environment.
- Your recordMCP for admin and building, not an audit-grade record.Your recordMCP is the governance backbone across every app.
What makes Superblocks governed in your own cloud compared to Retool
One stateless agent. Inference you already paid for.
Self-hosting Superblocks is a single stateless Docker agent per environment, no database to migrate, with features delivered over the air. Route Clark's inference through your own Snowflake, Databricks, or Bedrock so the AI runs inside your boundary and burns down spend you already committed.
Why this lasts: the teams that self-host for governance get the AI-native experience in their own cloud, not a cloud-only version they wait for.
A hospitality tech firm collapsed six databases across three VPCs down to a stateless agent per environment.
Clark gets smarter every time your team builds
Clark builds a knowledge graph across every builder, prompt, and integration: your design system, your coding standards, and the real schema of Snowflake, Databricks, and Salesforce. When the 50th builder queries a warehouse, they inherit what the first 49 already taught Clark, so the AI compounds instead of starting cold each session.
Why this lasts: reusable components are configured by a person. A knowledge graph is learned by the platform and grows with every build.
A large enterprise software company built across 100,000+ tables and 6,800 schemas, with Clark learning the estate so each builder builds with compounding knowledge.
A queryable system of record over every app
The Superblocks MCP exposes audit aggregation, full Clark chat history, and role management as tools, so you can answer who built what, on what data, and with whose approval in a single call. Policy agents run your own checks when an app publishes and block the ones that fail.
Why this lasts: querying and governing the whole estate from one record is an architecture, not an admin feature.
NHS Royal Surrey replaced Workday HR across the trust, GDPR compliant, scaling toward 100,000+ users.
Teams run real production on Superblocks
- Built it themselves on Superblocks: a recent grad advising the C-suite within a week, 800+ agents in production, and an $850K vendor quote turned down.
- A designer with no coding background built the trust's most sophisticated app, 60 to 70 APIs, replacing Workday HR, GDPR compliant.
- Migrating 200 spreadsheet workflows to governed apps. Quarter-long projects collapsed to under an hour, on their own infrastructure.
Hospitality tech company (left Retool) Ran six databases across three VPCs to keep self-hosted Retool isolated. Switched to a single stateless agent per environment.
Already on Retool? Move over on your own timeline.
No forklift migration. Keep Retool running while you build the next thing on Superblocks, and the Superblocks MCP gives IT visibility into both.
Pick the app Retool was going to struggle with, usually one that touches Snowflake, Databricks, or Salesforce, and build it AI-native first.
Keep your production Retool apps where they are. Add Superblocks for new builds. Govern both surfaces from one system of record.
When there's a reason, an upgrade, an AI rewrite, or a governance hook Retool can't provide, bring the existing app over and land it under your controls.
Retool vs Superblocks, where it counts
Skim it. These are the differences that show up the day your AI strategy scales.
| At scale | Retool |  |
|---|---|---|
| Architecture |
New AI builder bolted alongside the classic drag-and-drop product
|
One AI-native platform, rebuilt on React and TypeScript
|
| AI-native builder in self-hosted |
Cloud-only beta, coming to self-hosted later
|
Runs in your VPC today via the stateless agent
|
| Run it without the vendor |
Some files locked; apps run on Retool's runtime
|
Export to standard React and host it anywhere
|
| AI inference |
Connect model providers; newest models land on cloud first
|
Route through your Bedrock, Snowflake, or Databricks commit
|
| Self-hosted footprint |
Multiple services plus Postgres per VPC
|
Single stateless Docker agent per environment
|
| Upgrades |
Database migrations and scheduled downtime
|
Over-the-air features, simple agent upgrades
|
| System of record |
MCP builds apps and administers users
|
MCP with audit aggregation, chat history, role mgmt
|
| Governance at publish |
Reviewed separately
|
Policy agents check and block before an app ships
|
| AI that learns your data |
Reusable components you configure by hand
|
Knowledge graph learned across every builder and query
|
| Bring your own AWS |
Self-hosts in AWS, but the AI builder stays cloud-only
|
Build AI-native in your AWS account, inference through Bedrock
|
In one working session we deploy the stateless agent into your VPC, connect to your data, stand up audit logs and the Superblocks MCP against your identity provider, and have Clark build a real app against your real schema in the same hour. No script.
