idea×app

What can AI do today — and where it's heading

The tools that turn an idea into software are changing monthly, and most people only learn what they can do by seeing it. This page is our plain-language map: what actually works today, what is announced with dates, and what looks likely within a year — so you know what to prototype now and how to hand it off.

Last checked September 2026. Labels: Available now Announced & dated Likely Speculative. Predictions are our reading of public roadmaps and analyst reports, not promises.

What changed in the last few weeks

Why it feels impossible to keep up: 11 new models from 7 providers shipped in the first 11 days of September alone. Dated, with sources.

  1. OpenAI launches GPT-6 Astra with computer use

    It "can zip through spreadsheets, fill out forms, and navigate across web pages" on its own; Greg Brockman called it "not unreasonable" to say we are in the AGI era. Within a week people had rebuilt Manhattan in Unreal Engine, Apple Park in Blender and a 12-player browser shooter in a day — while Kotaku pointed out the game knock-offs are still "functionally worse". Both are true, and that is the gap this site is for. Source ↗

  2. Gemini 3.8 Flash ships — the third Flash release in six weeks

    Trackers count 11 new models from 7 providers in the first 11 days of September. Nobody can evaluate that many by reading release notes; you learn by trying prompts. Source ↗

  3. Claude Fable 5.1 arrives ~25% cheaper for typical workloads

    Anthropic's new default in Claude Code; Mythos 5.1 is the trusted-access version of the same model. Source ↗

  4. Pew: 34% of US adults have used AI chatbots for health questions

    A quarter used one to diagnose symptoms. Everyday use is far ahead of everyday understanding. Source ↗

  5. "The biggest gap isn't in the technology, it's in the understanding of it"

    Holden Richardson's state-of-AI note: nearly everyone knows the models have capabilities they don't understand, and no idea what those capabilities are. Showing working examples with their prompts is the fastest fix we know. Source ↗

  6. Gallup: 52% of US workers now use AI at work, 15% daily

    Organisational adoption jumped six points in one quarter — the sharpest rise Gallup has recorded. Source ↗

  7. Pew: 49% of US adults use chatbots; 63% say AI is advancing too fast

    Up from 33% in 2024. The "too fast" majority is exactly who a what-you-see-is-what-you-get prompt library is for. Source ↗

  8. Stanford AI Index 2026: generative AI reached 53% global adoption in three years

    Faster than the PC or the internet; agents went from 12% to 66% on the OSWorld computer-use benchmark in a year. Source ↗

Worth watching: People are Creating INSANE Things with GPT-6 Astra (RandomAI, ~7 min)

A good example of how people actually learn what is possible: a roundup of first-week builds — Manhattan in Unreal, 3D games through Blender, a Minecraft clone with natural disasters. Watch it, then try the same kind of thing here with your own prompt.

Now (September 2026)

Available today. Every item was checked against official docs or first-hand reports.

Coding agentsAvailable now

One prompt in a folder builds the first version

Claude Code, Codex (inside the ChatGPT desktop app, default model GPT-6 Astra), Cursor Agent, GitHub Copilot, Google Antigravity, Devin, Kiro and OpenCode all take a folder, read a spec, plan, write code, run tests and iterate. Claude Code is the most-used agent among professional developers (39% worldwide in JetBrains' 2026 survey).

Tools: Claude Code (Fable 5.1), Codex (GPT-6 Astra), Cursor, Copilot, Antigravity, Devin, Kiro, OpenCode

Source ↗

Desktop controlAvailable now

Agents can now use your computer, not just your code

The Codex app has a "Computer use" plugin that sees and operates macOS/Windows apps (ChatGPT Work or Codex plan; Windows is foreground-only). Claude Code's desktop Code tab offers Computer Use on Pro/Max (Mac/Windows) with view-only, click-only or full-control tiers per app. Google's Gemini Computer Use API (preview) drives browsers, Android and desktops with Gemini 3.8 Flash; Antigravity runs browser-in-the-loop agents for free.

Tools: Codex computer use, Claude Code Computer Use, Gemini Computer Use, Antigravity, Claude Cowork, Claude for Chrome

Source ↗

StandardsAvailable now

AGENTS.md and MCP are the shared plumbing

AGENTS.md (Linux Foundation's Agentic AI Foundation) is read natively by Codex, Cursor, Copilot, Antigravity, Kiro, OpenCode, Devin, Zed, Warp and Junie; Claude Code reads CLAUDE.md, which can import it. The Model Context Protocol, now also under the AAIF, connects agents to 10,000+ production servers — including game engines and design tools. That is why our build kit ships as AGENTS.md and lists MCP servers.

Tools: agents.md, MCP

Source ↗

Engines & 3DAvailable now

Agents work inside Unity, Blender and Godot

Unity ships an official AI Assistant and MCP Server (Unity 6, beta) that inspect GameObjects and drive editor actions; the community "MCP for Unity" adds 47 tools. blender-mcp (28k stars) lets Claude Code, Codex, Antigravity or Cursor build scenes, run Python, pull Poly Haven assets and generate meshes with Rodin or Hunyuan3D. godot-mcp creates scenes and nodes and runs projects. Unreal has community, experimental MCP servers only.

Tools: Unity AI Assistant + MCP Server, CoplayDev/unity-mcp, ahujasid/blender-mcp, @coding-solo/godot-mcp, chongdashu/unreal-mcp

Source ↗

Design toolsAvailable now

Design apps are agent-addressable

Figma's MCP server reads design context and, in beta, writes to the canvas; Figma Make builds from prompts. Webflow and Canva expose MCP servers, Framer's agent builds pages and components, Rive Agent is free for all users, Adobe Firefly has an AI Assistant that picks tools across image, video and audio.

Tools: Figma MCP (mcp.figma.com/mcp), Webflow MCP, Canva MCP, Framer AI, Rive Agent, Adobe Firefly

Source ↗

Generative mediaAvailable now

Images, video, 3D and music are cheap enough to prototype with

Nano Banana 2 images cost about $0.045 each, Veo 3.1 Lite video $0.05 per second, Meshy has a free 3D tier with an MCP server, Suno Pro is $8 a month, and Google's Genie 3 generates playable worlds at 720p for AI Ultra members. Prototype assets no longer need a budget line.

Tools: Nano Banana 2, Veo 3.1, Kling 3.0, Runway Gen-4.5, Meshy, Rodin, Hunyuan3D, Genie 3, Suno v6, ElevenLabs v3

Source ↗

ModelsAvailable now

Frontier models with 1M-token context are the default

Claude Fable 5.1 (1 Sep), GPT-6 Astra (3 Sep) and Gemini 3.8 Flash (2 Sep) all shipped this month. Agents now read whole codebases and long specs at once — a single HTML prototype fits comfortably.

Tools: Claude Fable 5.1, GPT-6 Astra, Gemini 3.8 Flash

Source ↗

Next 6 months (to March 2027)

Announced with dates, or the obvious next step of something already shipping.

PhonesAnnounced & dated

Siri acts inside apps on iOS 27 (14 Sep 2026)

iOS 27 ships Siri AI with on-screen awareness and in-app actions through App Intents, running on-device or in Private Cloud Compute. Expect "the assistant does it for you" to become a normal expectation of phone apps — worth designing for in prototypes.

Tools: Siri AI, App Intents

Source ↗

VideoAnnounced & dated

Sora API shuts down 24 Sep 2026

OpenAI's Sora app closed in April and the API closes this month; Veo, Kling and Runway are the remaining mainstream video generators. If a prototype mentions video, plan around those.

Tools: Veo 3.1, Kling 3.0, Runway

Source ↗

PricingAnnounced & dated

Gemini 3.x introductory prices end 31 Dec 2026

Flash-class output prices roughly double after the promo. Free tiers stay, but heavy users of this site should budget for it — or lean on Flash-Lite, which stays cheap.

Tools: Gemini 3.8 Flash, 3.5 Flash-Lite

Source ↗

CursorAnnounced & dated

Multi-agent "Projects" coordinate cloud subagents

Cursor's Projects (10 Sep) put a coordinator over cloud subagents; Codex added worktrees and a Windows background daemon the same week. One brief, several agents, parallel milestones is becoming the default shape of a build.

Tools: Cursor Projects, Codex CLI 0.154

Source ↗

DesignAnnounced & dated

Figma's write-to-canvas MCP leaves beta and becomes usage-billed

Agents drawing directly into Figma is free during beta; Figma has said it will be metered afterwards. Design handoff from a prototype to a real design file is about to be routine — and paid.

Tools: Figma MCP

Source ↗

ModelsLikely

Gemini 3.5 Pro lands; Claude Haiku 4.5 retires

Gemini 3.5 Pro was announced at I/O 2026 and is "pending testing"; Anthropic retires Claude Haiku 4.5 on or after 15 Oct 2026. Expect OpenAI DevDay and Adobe MAX in October to push GPT-6 agent tooling and Firefly agents further.

Tools: Gemini 3.5 Pro, Claude Haiku 4.5

Source ↗

Desktop controlLikely

Computer use reaches cheaper plans and Windows agents exit preview

Anthropic and OpenAI both have desktop control in production on Mac and Windows behind top-tier plans; Microsoft's on-device agent registry, File Explorer and Settings MCP connectors are in prerelease. The pattern usually broadens within two quarters.

Tools: Claude Code, Codex, Windows agent platform

Source ↗

Within 12 months (to September 2027)

Our reading of public roadmaps and analyst reports — plausible, not promised.

EnginesLikely

Official engine MCP servers become the norm

Unity already ships one; the community versions for Blender, Godot and Unreal are the templates vendors are likely to adopt. A game prototype plus a brief will increasingly go straight into an engine scene built by an agent.

Tools: Unity, Godot, Blender, Unreal

Source ↗

On-deviceLikely

Small local models handle UI actions

Apple's on-device foundation models (3B and 20B), Windows Foundry Local and Gemma 4 point to local sub-agents that tap, type and file things without the cloud — cheaper, private, always on.

Tools: Apple Foundation Models, Foundry Local, Gemma 4

Source ↗

EconomicsLikely

Prices keep falling 25–45% per generation; 1M context is table stakes

Fable 5.1 arrived about 25% cheaper than its predecessor and GPT-5.6 Luna costs $0.20 per million input tokens. The cost of trying an idea keeps heading toward zero, which shifts the scarce thing to knowing what to ask for.

Source ↗

Reality checkAvailable now

Most "agents" in production still aren't, and AI code needs review

Menlo Ventures counted only 16% of enterprise generative-AI deployments as true agents, and independent analyses of AI co-authored pull requests find markedly more major issues than human ones. Faster building raises the value of validating the idea first — which is what a prototype is for.

Source ↗

WorldsSpeculative

Prompt-to-playable worlds that export to engines

Genie 3, Tencent's HY-World 2.0 and "WorldClaw" already generate explorable worlds from text; the open question is whether they export clean, editable scenes. If they do, a one-sentence game idea could become an engine project, not just a browser demo.

Tools: Genie 3, HY-World 2.0, Hunyuan3D-WorldClaw

Source ↗

AgentsSpeculative

An agent builds a full Unity or Unreal scene from a spec, end to end

OpenAI's own GPT-6 Astra examples include building video-game scenes; combined with engine MCP servers and desktop control, a brief like the ones this site produces could drive an entire scene build. Plausible within a year; not a product today.

Tools: GPT-6 Astra, Unity MCP, Codex computer use

Source ↗

How a prototype-first workflow fits

Prototype first, then hand off

A clickable prototype is the cheapest spec there is: people react to it in seconds, and every coding agent can read it as the reference for the real build.

AGENTS.md from day one

Our build kit writes the brief in the file agents already read, lists the MCP servers for engines and design tools, and adds a CLAUDE.md pointer — nothing is lost between "I have an idea" and "the agent is building it".

Your key, your tools

Generation runs with your own Gemini key; the kit works with whichever agent you already pay for. No lock-in in either direction.

Games included

2D and 3D game prototypes come out playable, and their kits point at Phaser or Three.js — and at the Unity, Godot and Blender integrations agents can drive today.

Validation matters more, not less

When building is cheap, knowing what to build is the scarce part. A prototype with real reactions and comments tells an agent what to make; that is the step before the agent loop, not a rival to it.

Prototype an idea now See what people built

Beyond code: agents that use creative tools

Coding agents no longer stop at code. Through MCP servers and native integrations they can open a game engine, a 3D tool or a design file and work inside it — and some can operate your whole desktop. Here is what is real today, with the exact names to search for.

Unity

Official Unity AI Assistant and MCP Server (Unity 6, beta; MCP free, Assistant credits from $10/month) inspect GameObjects, drive editor actions and verify results. The community "MCP for Unity" (CoplayDev/unity-mcp) adds 47 tools for scenes, scripts, assets, tests and builds.

How: Install the Unity MCP Server package or CoplayDev/unity-mcp, connect Claude Code, Codex, Cursor or Antigravity, then hand over AGENTS.md and the prototype.

Blender

blender-mcp (ahujasid, 28k stars) lets an agent build scenes and materials, run Python, pull Poly Haven and Sketchfab assets, and generate meshes with Hyper3D Rodin or Hunyuan3D. Blender itself (5.2) ships no built-in AI.

How: Add the blender-mcp add-on and MCP server; works with Claude Code, Codex, Antigravity and Cursor.

Godot

godot-mcp (@coding-solo/godot-mcp) launches the editor, runs projects, captures debug output and creates scenes, nodes and sprites for Godot 4.4+.

How: npx-installable MCP server; point your agent at it and at the prototype.

Unreal Engine

UE 5.8 has no official MCP server yet; chongdashu/unreal-mcp (experimental) can create actors, Blueprint classes and graph nodes and control the viewport.

How: Community server, expect rough edges — good for spikes, not production pipelines.

Figma

Figma's official MCP server reads design context, variables and components, generates code, and — in beta — writes to the canvas; Figma Make builds from prompts and connects to MCP.

How: Remote server at mcp.figma.com/mcp on any plan; desktop server with Dev/Full seats.

Webflow · Framer · Canva · Rive

Webflow's MCP server edits elements, styles, components, CMS and SEO; Framer's AI agent builds pages and code components and connects to Claude Code, Cursor and Codex; Canva offers a Dev MCP server and AI Connector; Rive Agent is free for all users.

How: Each vendor documents an MCP or connector setup; add it to your agent's MCP config.

Desktop control

Codex "Computer use" (ChatGPT Work/Codex plans, Mac + Windows, foreground-only on Windows), Claude Code desktop Computer Use (Pro/Max, Mac + Windows, per-app permission tiers), Claude Cowork and Claude for Chrome, Google Antigravity browser agents (free), Gemini Computer Use API (preview), Microsoft Copilot Studio computer use (enterprise, registered Windows machine).

How: Give the agent the prototype and the brief, then let it drive the tool you already use — but keep the permissions narrow and review what it does.

Generative media

Images: Nano Banana 2 (~$0.045), GPT Image, FLUX 3, Midjourney V8. Video: Veo 3.1, Kling 3.0, Runway Gen-4.5 (Sora API ends 24 Sep 2026). 3D: Meshy (free tier, MCP server), Rodin, Hunyuan3D. Music and voice: Suno v6, ElevenLabs v3, Lyria 3.5.

How: Mention the assets you need in the brief; most of these have APIs or MCP servers an agent can call.

Your Gemini API key

Generation runs in your browser with your own free Google Gemini key. The key is saved only on this device (browser storage) and is sent only to Google — never to this site.

  1. Open Google AI Studio → API keys and sign in with a Google account.
  2. Click Create API key and copy it.
  3. Paste it below and press Save.