We built InsForge because we wanted to use pure prompts to build a production-grade app faster. In practice, AI isn’t reliable enough when generating code or configuring backend pieces like schema, RLS, auth, or functions. Postgres MCP and Supabase MCP are a good starting point, but they felt more like API wrappers for agents than something designed to guide AI safely through backend tasks.

So we built our own MCP server on top of Postgres and added a set of context-engineering tools to make prompt-driven workflows more predictable. It gradually turned into a Postgres-based BaaS.

Key features:

1. Authentication with prebuilt UI components

2. Postgres database with typed SDK

3. Serverless functions with a secret manager

4. S3-compatible file storage

5. AI model integration with a unified inference API and simple GUI

6. MCP server with context-engineering endpoints (`fetch-docs`, `get-backend-metadata`, `get-table-schema`, etc.)

We're still early (4 months in). Some parts are rough, but we’re shipping daily and improving quickly :)

[1] Launch blog: https://insforge.dev/blog/insforge-launch

[2] Open-source repo (self-hosting): https://github.com/InsForge/InsForge

[3] Benchmark: https://github.com/InsForge/mcpmark

[4] Website (cloud hosting): https://insforge.dev/

[5] Our article on context engineering: https://insforge.dev/blog/why-context-is-everything-in-ai-co...

[6] Public roadmap: https://feedback.insforge.dev/roadmap