PromptAPI Runtime

PromptAPI Runtime is a high-level JavaScript/TypeScript framework for building an LLM-as-API server from a user-supplied systemPrompt.

The framework exposes one server endpoint:

POST /api/chat

The user sends:

• systemPrompt: the endpoint registry and API behavior contract
• provider: gemini, openai, anthropic, mistral, or openai-compatible
• model: provider model name
• message: natural-language API request

The runtime lets the model call tools, especially Prisma-backed database tools, then returns a JSON response with status_code.

---

Architecture

User / Client
  ↓
POST /api/chat
  ↓
PromptAPI Runtime
  ↓
Provider Adapter: Gemini / OpenAI / Anthropic / Mistral / OpenAI-compatible
  ↓
Tool Calling
  ↓
Prisma Database Tools
  ↓
Strict JSON API-style response

The system prompt acts as the server contract:

• endpoint registry
• validation rules
• required fields
• status code behavior
• database tool behavior
• daily testing behavior

The backend remains the execution authority:

• Prisma saves and retrieves records
• tools execute outside the model
• audit logs are persisted
• daily tests are stored

---

Supported providers

Provider	Adapter	Env var
Google Gemini	gemini	GEMINI_API_KEY
OpenAI	openai	OPENAI_API_KEY
Anthropic Claude	anthropic	ANTHROPIC_API_KEY
Mistral	mistral	MISTRAL_API_KEY
OpenAI-compatible providers	openai-compatible	OPENAI_COMPATIBLE_API_KEY, OPENAI_COMPATIBLE_BASE_URL

Use openai-compatible for providers that expose OpenAI-compatible chat completions with tools.

---

Install

npm install
cp .env.example .env

Fill your .env:

DATABASE_URL="file:./dev.db"
DEFAULT_PROVIDER=gemini
DEFAULT_MODEL=gemini-2.5-flash
GEMINI_API_KEY=your_key_here

Initialize Prisma:

npm run prisma:generate
npm run prisma:migrate

Start the server:

npm run dev

---

One endpoint usage

curl -X POST http://localhost:3000/api/chat \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "gemini",
    "model": "gemini-2.5-flash",
    "message": "Create customer c_100 named Loai Abdalslam with email loaiabdalslam@gmail.com"
  }'

Example response:

{
  "status_code": 201,
  "endpoint": "create_customer",
  "method": "POST",
  "request_id": "req_...",
  "provider": "gemini",
  "model": "gemini-2.5-flash",
  "message": "Customer created successfully",
  "data": {
    "customer_id": "c_100",
    "name": "Loai Abdalslam",
    "email": "loaiabdalslam@gmail.com"
  },
  "errors": [],
  "meta": {
    "tool_calls_count": 1
  }
}

---

Request schema

type ChatRequest = {
  message: string;
  systemPrompt?: string;
  provider?: "openai" | "gemini" | "anthropic" | "mistral" | "openai-compatible";
  model?: string;
  temperature?: number;
  mode?: "chat" | "daily_test";
  metadata?: Record<string, unknown>;
};

---

Built-in Prisma tools

The model can call these tools:

db_create_record(namespace, key, value)
db_get_record(namespace, key)
db_search_records(namespace, query?, limit?)
db_update_record(namespace, key, patch)
db_delete_record(namespace, key)

These tools are intentionally generic, so any system prompt can define its own logical endpoints over the same persistence layer.

---

Daily endpoint tests

Run from HTTP:

curl -X POST http://localhost:3000/api/chat \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "gemini",
    "model": "gemini-2.5-flash",
    "mode": "daily_test",
    "message": "run daily endpoint tests"
  }'

Or from CLI:

npm run daily:test

Daily test results are saved to Prisma model DailyTestRun.

---

Custom system prompt

Send your own system prompt in the request:

{
  "provider": "gemini",
  "model": "gemini-2.5-flash",
  "systemPrompt": "You are my API server... Endpoints: ... Tools: ...",
  "message": "Create supplier medicine..."
}

See:

examples/custom-system.prompt.txt

---

Recommended system prompt structure

You are a System-Prompt-Defined API Server.
Return strict JSON only.

Global response schema:
{ status_code, endpoint, method, message, data, errors, meta }

Tools:
- db_create_record
- db_get_record
- db_search_records
- db_update_record
- db_delete_record

Endpoints:
1. create_customer
Method: POST
Required: customer_id, name
Database: Save namespace customers using customer_id as key.
Success: 201

Rules:
- Use tools before final answer.
- Never claim saved unless tool result ok=true.

---

Notes

• The system prompt is not literally a network server. It is the API contract and routing brain.
• The Node.js runtime is the actual server.
• The model selects endpoints and calls tools.
• Prisma performs database execution.
• The framework returns API-style JSON responses.

---

Production checklist

Before production, add:

• authentication
• tenant isolation
• tool permission policies
• rate limiting
• request signing
• stronger JSON-schema validation per endpoint
• provider retry logic
• observability dashboards
• prompt injection hardening
• backup database, e.g. PostgreSQL instead of SQLite

---

Citation

Abdalslam, L. (2026). LLM as API Server: A System-Prompt-Defined Runtime for Natural-Language Endpoint Routing, Tool Calling, and Database-Backed Execution. SSRN Working Paper. https://ssrn.com/abstract=...

License

MIT