Prompt-driven backend framework
Replace REST endpoints with natural language. Input is NL, processing is LLM reasoning + tool orchestration, output is structured response. A controlled hallucination engine with guardrails.
9 pre-registered tools. All actions go through strict schema validation.
Transform data with 6 operations — uppercase, lowercase, json_parse, json_stringify, length, sum.
Create a database record with owner scoping. Content stored as JSON with automatic type and metadata.
Search records by type with pagination. Results limited to 100 max. Owner-scoped by default.
Update an existing record by ID. Validates ownership before mutation. Returns updated data.
Delete a record by ID with ownership verification. Destructive action — gated behind auth check.
Verify API keys via SHA-256 hash lookup. Returns user identity on success. Entry point for authenticated flows.
Look up user by ID. Self-access only unless admin. Defense-in-depth with dual permission check.
List all users with pagination. Admin only. Passwords excluded from output. Limit capped at 100.
System statistics — user count, record count, DB type. No authentication required. Dashboard feed.
pip install vibeserver
vibeserver init my-app
vibeserver serve
POST /
{"input": "greet Alice"}
→ {"message": "Hello, Alice!"}
Persistent NL→plan cache. No embedding model. Repeats skip the LLM entirely.
19 injection patterns
Role-based blocking (banned)
Output key sanitization
Tool Registry schema validation
JSonschema Draft7 compliance
System prompt hardening