Dropbase: AI-generated Python internal tools builder; local-first, self-hosted

Dropbase is a self-hosted platform for building internal tools — admin panels, billing dashboards, ops tooling — by generating Python + JSON app code from natural-language prompts. The generated code is yours: inspect it, edit it, commit it. Local-first by design: credentials and data stay on your infrastructure.

Two-file architecture

A Dropbase app consists of two files: properties.json (declarative UI — components, layouts, data bindings) and main.py (backend logic — database queries, third-party API calls, business rules). The AI generates both from a prompt.

The constraint is deliberate. Because the structure is deterministic and minimal, frontier LLMs can reliably regenerate the complete app from a new prompt — rather than attempting surgical edits to an arbitrary file tree. The creator confirmed in the Show HN thread that the team chose full-file regeneration over AST-level editing specifically because it produces more consistent results.

Self-hosted deployment

Run the server locally with Docker via start.sh. AI generation requires an API key for OpenAI or Anthropic configured in server.toml on your machine. Database credentials and third-party tokens (Salesforce, HubSpot, Slack, Mailgun) go in worker.toml in your workspace directory. Nothing is sent to Dropbase's servers unless you opt into cloud features.

Third-party integrations beyond the pre-wired set are imported as standard PyPI packages inside main.py — the full Python ecosystem is available.

LLM requirement: frontier class only

Dropbase requires GPT-4o or Claude 3.5 Sonnet for consistent results. From the Show HN thread (item 40947415, July 2024), creator Omar stated: "we could not achieve the consistent results we needed with any LLM prior to these models." Teams running open-source or cost-reduced models should expect unreliable code generation. The two-file constraint reduces the problem surface but does not compensate for weaker reasoning.

What the HN community raised at Show HN (July 2024)

The Show HN thread (141 points, July 12, 2024, item 40947415) surfaced three concrete observations from developers evaluating the platform.

Code-regeneration vs. AST-level editing. A commenter building similar technology asked how Dropbase handles modification reliability: does it use ASTs for targeted edits or regenerate files wholesale? The creator confirmed: both properties.json and main.py are regenerated entirely on each AI interaction, with prompt engineering preserving existing functionality. The commenter noted their own approach uses ASTs for safer edits; the creator acknowledged the scalability concern with larger files but confirmed the regeneration approach was intentional.

Framework lock-in. A commenter asked whether Dropbase supports existing Python web frameworks — specifically Django or Litestar. The answer is no: Dropbase uses its own proprietary framework. Existing Python code can be imported into main.py as a library, but the app container is Dropbase-native. Teams migrating an existing Flask or Django service would rewrite rather than migrate.

Audience beyond AI-enthusiasts. One commenter noted that "writing two files to build a webapp" could appeal to developers frustrated with Retool's pricing or Appsmith's complexity — teams that adopt Dropbase for the Python-native workflow, not the AI generation specifically. The AI lowers the floor on getting to a working first version; the code-first output raises the ceiling for customization beyond what drag-and-drop builders allow.

Sources: community-thread — HN item 40947415 (2024-07-12, Show HN: Dropbase AI – A Prompt-Based Python Web App Builder, 141 points); ProductHunt launch — dropbase-ai (2024-06-20, 181 upvotes); GitHub — github.com/DropbaseHQ/dropbase (1.3k stars, observed 2026-05-20).