Your Backend Is Now a Prompt and a JSON Blob

What it does

backend-GPT is a JavaScript experiment that collapses the entire backend and database into a single LLM prompt. You describe the app’s purpose in plain text, seed a JSON state blob of up to about a kilobyte, and start firing off pseudo-API calls. The model infers what each call means, mutates the state, and returns a response, eliminating the need for explicit routes or schemas.

The interesting bit

The project treats API names as loose suggestions rather than rigid contracts: mistype an endpoint or ask for get_board_state_as_fen() instead of get_board_state(), and the LLM is expected to adapt. The authors pitch this as “serverless without cold starts,” since the only difference between tenants is a 1 KB state blob and a few lines of instruction.

Key highlights

• Infers CRUD logic from function-like or natural-language commands (e.g., add_five_housework_todos(), delete last todos)
• Persists state as a single JSON blob capped at roughly one kilobyte
• Swaps contexts by swapping prompts and state, with no deployed code to warm up
• Built in JavaScript and currently sitting at 2,933 GitHub stars

Caveats

• State is explicitly limited to about a kilobyte, so this is strictly toy-scale
• Natural-language instructions work, but the README notes the model “tends to work better” with pseudo-function syntax
• How the LLM is hosted, latency, and running costs are left unexplained in the README

Verdict

Good for frontend prototypes and LLM stunts; bad for production workloads that exceed a kilobyte or require the operational specifics the README omits.