Backends & Guarantee Levels

Backend	Guarantee	Mechanism
openai()	native	Server-side strict JSON schema
groq()	native	JSON mode
deepseek()	native	JSON mode (no schema-strict mode, same tier as groq())
fireworks()	native	Server-side JSON schema mode, plus a real token-level GBNF grammar mode
mistral()	native	Server-side JSON schema mode
gemini()	native	Server-side JSON schema mode (responseJsonSchema)
ollama()	constrained	Token-level JSON-schema constraint
llamaCpp()	constrained	Token-level GBNF grammar (local .gguf via node-llama-cpp)
anthropic()	best-effort	Prompt + parse + retry
openRouter()	best-effort	Pass-through to many providers - response_format support varies by underlying model

llamaCpp() is constrained for a { gbnf } input (token-level). For other schema types (Zod / jsonSchema / …) it currently runs a best-effort prompt path until the JSON-Schema→GBNF converter lands - treat those as best-effort despite the nominal level.

fireworks() is the one cloud backend where a { gbnf } input is not downgraded to best-effort - Fireworks' grammar mode (response_format: { type: "grammar", grammar }) applies the GBNF grammar as a genuine token-level constraint server-side, the same guarantee llamaCpp() gives locally. It reuses the openai package pointed at Fireworks' base URL, so no extra SDK dependency is needed.

openRouter() is deliberately best-effort, not native like the other cloud backends - it's pass-through across many different underlying providers/models, and response_format: { type: "json_schema" } enforcement isn't guaranteed for every model it can route to, only the ones that actually support it themselves.

gemini() uses the official @google/genai SDK, not an OpenAI-compatible endpoint (unlike fireworks()/mistral()/openRouter()) - Gemini's OpenAI-compat layer is a migration bridge for OpenAI users, not its primary integration path, and doesn't expose responseJsonSchema (plain JSON Schema, what toJsonSchema() already produces) - only the older responseSchema (Gemini's own Type-enum OpenAPI-subset shape).

import { openai, groq, fireworks, mistral, gemini, openRouter, deepseek, ollama, anthropic, llamaCpp } from "@aviasole/shapecraft";

const gpt       = openai({ model: "gpt-4o-mini" });
const fast      = groq({ model: "llama-3.3-70b-versatile" });
const cloudGbnf = fireworks({ model: "accounts/fireworks/models/llama-v3p1-70b-instruct" });
const mist      = mistral({ model: "mistral-large-latest" });
const gem       = gemini({ model: "gemini-flash-latest" });
const router    = openRouter({ model: "openai/gpt-4o-mini" });
const deep      = deepseek({ model: "deepseek-v4-flash" });
const local     = ollama({ model: "llama3.2" });
const native    = llamaCpp({ modelPath: "./models/llama-3.2-3b.gguf" });
const claude    = anthropic({ model: "claude-haiku-4-5-20251001", maxRetries: 3 });

Model Capabilities

Every built-in backend also exposes capabilities - an explicit, inspectable alternative to duck-typing typeof model.generateStream === "function" for routing logic:

console.log(claude.capabilities);
// { streaming: true, chat: true, structuredOutput: true, toolCalling: true, skillDispatch: true }

interface ModelCapabilities {
  streaming: boolean;        // has generateStream()
  chat: boolean;             // has chat() - required for turnaround: true
  structuredOutput: boolean; // has generate() - always true
  toolCalling: boolean;      // native provider function-calling, drives generateWithTools() -
                              // true wherever the backend implements toolCall(); llamaCpp() is
                              // the one backend without it (local GGUF exposes no tools API)
  skillDispatch: boolean;    // generateSkillCall()/runSkillLoop() - always true, built on generate()
}

capabilities is optional on ShapecraftModel - a custom model implementation that predates this field (or simply doesn't set it) still satisfies the interface unchanged, and model.capabilities is undefined for it. chat?/generateStream? remain the actual methods the core calls; capabilities is just a declared summary of the same information, not a replacement mechanism.

See Tool Calling for toolCalling/generateWithTools(), and Skill-Based Generation for skillDispatch.