AI Models

ForgeAI integrates with OpenAI through two distinct paths: the @inngest/agent-kit framework for the primary coding agent network, and the OpenAI SDK directly for vision-based design spec extraction.

Required API keys

Variable	Description
`OPENAI_API_KEY`	Required for all OpenAI model calls
`ANTHROPIC_API_KEY`	Available for alternative model routing
`GEMINI_API_KEY`	Available for alternative model routing

ForgeAI is built on @inngest/agent-kit, which abstracts model calls through a unified agent and network API. You can swap models per agent without changing tool or network logic.

Models in use

Coding agent (GPT-5.2)
Design spec extraction (GPT-4o)
Naming agent (GPT-5.2)

The primary agent that generates, reads, and updates project files inside the E2B sandbox.

inngest/functions.ts

import { createAgent, openai } from "@inngest/agent-kit";

const codeAgent = createAgent({
  name: "coding agent",
  system: PROMPT,
  description: "An expert coding agent",
  model: openai({
    model: "gpt-5.2",
  }),
  tools: [terminal, createOrUpdateFiles, readFiles, unsplashImage],
});

The agent runs inside a createNetwork with maxIter: 20. The network router exits when the agent sets a <task_summary> in its response, which is detected via the onResponse lifecycle hook.

inngest/functions.ts

const network = createNetwork({
  name: "codeing-agent-network",
  agents: [codeAgent],
  maxIter: 20,
  defaultState: codingAgentState,
  router: async ({ network }) => {
    if (network.state.data.summary) {
      return; // exit network
    }
    return codeAgent;
  },
});

When a user uploads a screenshot or design image, GPT-4o-2024-08-06 parses it into a structured DesignSpec object using OpenAI’s structured output feature with zodResponseFormat.

lib/extract-design-spec.ts

import { zodResponseFormat } from "openai/helpers/zod";
import OpenAI from "openai";

const openai = new OpenAI();

const response = await openai.chat.completions.parse({
  model: "gpt-4o-2024-08-06",
  messages: [
    { role: "system", content: DESIGN_PROMPT },
    {
      role: "user",
      content: [
        { type: "text", text: `User hint: ${userHint}` },
        { type: "image_url", image_url: { url: imageUrl } },
      ],
    },
  ],
  response_format: zodResponseFormat(DesignSpecSchema, "design_spec"),
  temperature: 0.1,
});

The parsed spec includes theme colors, page routes, section layouts, and content blocks. It is serialized as JSON and prepended to the coding agent’s input prompt.

After code generation completes, a lightweight naming agent generates a short, product-like project title.

inngest/functions.ts

const metadataAgent = createAgent({
  name: "Forge ai project namer",
  system:
    "You create short, product-like project names (5 - 10 words, Title Case). Return ONLY the name",
  model: openai({
    model: "gpt-5.2",
  }),
});

The agent receives the original user prompt and returns a plain-text title (no quotes, 5–10 words, Title Case, max 80 characters). The result is saved directly to the project record in the database.

Agent network state

The coding agent network tracks two pieces of shared state across iterations:

inngest/functions.ts

interface CodeAgentState {
  summary: string;   // set when the agent includes <task_summary> in its response
  files: Record<string, string>; // accumulated file path → content map
}

State is initialised with previous code files and message history loaded from the database, giving the agent context from prior runs on the same project.

Agent-kit primitives used

Primitive	Purpose
`createAgent`	Define an agent with a model, system prompt, and tools
`createNetwork`	Compose agents into a routed loop with shared state
`createTool`	Define a typed tool with a Zod parameter schema and async handler
`createState`	Initialise typed state shared across network iterations

Inngest background jobs

How the agent network runs as a durable background function.

Design to code

How uploaded screenshots are converted to design specs and forwarded to the agent.

Get Started

Core Features

Configuration

Integrations

Required API keys

Models in use

Agent network state

Agent-kit primitives used

Inngest background jobs

Design to code

Build docs developers (and LLMs) love

Get Started

Core Features

Configuration

Integrations

​Required API keys

​Models in use

​Agent network state

​Agent-kit primitives used

​Related pages

Inngest background jobs

Design to code

Build docs developers (and LLMs) love

Required API keys

Models in use

Agent network state

Agent-kit primitives used

Related pages