Structured output for LLMsReliable structured AI runtime for Node.js
Token-level constraints for local models, native JSON modes for cloud APIs - one unified API with explicit guarantee levels, not silent best-effort parsing.
import { z } from "zod";
import { generate, openai } from "@aviasole/shapecraft";
const Person = z.object({
name: z.string(),
age: z.number(),
});
const result = await generate(
openai({ model: "gpt-4o-mini" }),
Person,
"John Doe is 32"
);
result.data
// { name: "John Doe", age: 32 }
result.guaranteeLevel
// "native"npm install @aviasole/shapecraftGuarantee levels
openai()
Server-side strict JSON schema enforcement
groq()
JSON mode
deepseek()
JSON mode
fireworks()
Grammar-based cloud constraint
mistral()
JSON mode
gemini()
Server-side JSON schema mode
ollama()
GBNF grammar, token-level
node-llama-cpp()
GBNF grammar, token-level, fully local
anthropic()
Prompt + parse + retry
openRouter()
Pass-through, depends on underlying model
What else is in the box
Streaming
Incremental JSON parsing with partial-object events
createClient() middleware
Koa-style pipeline for logging, retry, timeout defaults
Batch generation
Concurrency-capped generateBatch() across many prompts
FHIR R4 presets
Patient, Observation, Condition, MedicationRequest, Encounter
Skill-based generation
Model dispatches to registered skills, agentic loop support
Tool calling
Native provider function-calling with validated arguments
Multi-agent orchestration
runAgents() chains validated generate() calls with your own router
Turnaround
Multi-turn collection, validated once over the whole transcript
CLI
npx shapecraft validate - check a JSON file against a schema, no code