"Vibe coding" is the current name for a way of building software that is spreading fast. You describe in plain language what you want to build, and the AI builds it. You steer the work and the AI writes the code.
This works well for frontend applications. It often breaks down on the backend.
The Backend Problem in Vibe Coding
Ask an AI to build you a React todo app and it does a fine job. Ask it to also set up the backend, with authentication, data persistence, permissions and real-time updates, and the work gets harder. The AI has to make assumptions about your infrastructure, generate server code, configure databases and handle deployment. Every wrong assumption means a debugging session that interrupts the work.
With Ulabase the backend already exists. You get a MongoDB backend with REST APIs, GraphQL, WebSocket, JWT auth, user management and role-based permissions in minutes, with no server code to write. The AI can spend the session on your product logic.
What Ulabase Gives the AI to Work With
When Claude Code (or any AI coding assistant) works with Ulabase, it has a defined API surface to generate against:
Data APIs: REST endpoints for every MongoDB collection. Create a collection and you get full CRUD with filtering, sorting, pagination, and aggregation, with no backend code.
Authentication: the /users and /acl collections manage users and permissions. The AI generates the Authorization headers and the permission documents from the documented format, without guessing.
GraphQL: define a schema and RESTHeart generates the resolvers. This is useful when the AI is building a complex frontend that needs flexible queries.
Real-time: WebSocket and Server-Sent Events for change streams. The AI can add live updates to any collection with a single endpoint.
Aggregations: expose MongoDB aggregation pipelines as HTTP endpoints. The AI writes the pipeline once, and from then on it runs as an API.
This surface is concrete and documented, so the AI knows what the backend offers and does not have to imagine it.
Claude Code + Sophia MCP
Ulabase runs Sophia, an AI knowledge base built on RESTHeart itself. Sophia exposes the complete Ulabase documentation via MCP (Model Context Protocol).
Claude Code can connect to it directly:
claude mcp add --transport http ulabase-docs https://d0c5a1.ulabase.app/mcp
Once connected, Claude Code can look up every Ulabase API, configuration option, and example. It reads the current documentation, retrieved at query time via semantic search, and does not depend on training data from a year ago.
Without Sophia, a vibe coding session goes like this:
"Build me a product catalog with search and filtering" → AI guesses at a backend structure → generates backend boilerplate → you debug it → you explain the actual API → repeat
With Sophia connected, it goes like this:
"Build me a product catalog with search and filtering using Ulabase" → Claude Code queries Sophia for how collections, ACL, and text search work → generates the exact
curlcommands to set up the backend → generates the frontend against the real API → done
The AI works from the docs, so it does not guess the API and you do not debug generic boilerplate.
A Concrete Session
A vibe coding session looks like this. You open Claude Code and say:
"I want to build a simple expense tracker. Users log in, add expenses with a category and amount, and can see their total by category. Use Ulabase for the backend and React for the frontend."
Claude Code, with Sophia MCP connected, does the following:
- Queries Sophia for how to set up collections, user authentication, and role-based permissions in Ulabase
- Generates the backend setup:
PUT /expensesto create the collection, a permission document that restricts each user to their own data (areadFilteron@user._id), and an aggregation endpoint for category totals - Generates the React frontend: login form using JWT auth, expense entry form, category breakdown chart, all calling the Ulabase API it just set up
- Runs everything, and you test it immediately
The entire backend runs in Ulabase. Neither you nor the AI wrote server code: the AI configured existing infrastructure via API calls.
Why This Combination Works
With a generic backend the AI has to invent the infrastructure and hope it works. Four properties of Ulabase help an AI coding session:
- The API is declarative. Creating a collection, setting permissions and defining an aggregation are data operations. An AI generates data well.
- The surface is small and consistent. Ulabase's API follows clear patterns. An AI that has read the docs can generate correct calls.
- Sophia keeps the AI current. With the MCP connection, Claude Code retrieves the current documentation at query time and does not work from stale training data.
- The feedback loop is short. There is no compilation and no deployment pipeline, so you test in seconds. When something is wrong, you fix the prompt.
Getting Started
- Create a Ulabase account at ulabase.com: the Free plan is enough to start
- Connect Claude Code to Sophia:
claude mcp add --transport http ulabase-docs https://d0c5a1.ulabase.app/mcp - Start a session: describe your application, and tell Claude Code you're using Ulabase
Ulabase is a MongoDB Technology Partner. Sophia is an AI knowledge base built on Ulabase: try it live.
Build your app on Ulabase
Your backend is ready in minutes, and you spend your time on your app. The Free plan needs no credit card.