Limitations¶
Documenting what Whittl does NOT do is as important as documenting what it does. If you know where the edges are, you can trust the middle more.
Platforms¶
- macOS is not supported. No build, no in-development macOS target. Whittl may work on Linux-in-a-VM on a Mac, but I can't help troubleshoot it.
- Windows 10 is the minimum. Windows 7 / 8 / 8.1 will not install.
- Linux — tested on Ubuntu 22.04+, Fedora 38+, Arch, Debian 12+. Older distributions may work but aren't tested.
- 32-bit systems are not supported. Both Whittl and its bundled Python runtime are 64-bit only.
Languages¶
Whittl generates Python. That's the whole ecosystem. It does not generate:
- JavaScript, TypeScript, Rust, Go, C, C++, Java, etc.
- HTML / CSS (unless your Python app outputs HTML, like a Flask template)
- SQL (beyond what a Python app needs to execute — no standalone schema designer)
If you need a non-Python target, Whittl isn't your tool.
GUI frameworks¶
Whittl's four targets cover most desktop and CLI use cases but not all:
- Kivy is not supported.
- wxPython is not supported.
- Web frontend (React, Vue, Svelte) is out of scope — Whittl ships local apps, not web apps.
- Game engines (Godot, Unity, Unreal) are not supported. Pygame works in the General Python target but there's no dedicated pipeline.
- iOS apps are not supported (Flet targets Android only in Whittl).
See Choosing a Target for what IS covered.
AI backends¶
- Only the five supported backends. Custom endpoints are limited to OpenAI-compatible via a configured base URL — no fully custom protocol support.
- No batched generation. One generation per turn. Want 10 variations of the same prompt? Run it 10 times.
- No fine-tuning pipeline. Whittl uses base models from each backend. You can't upload a custom-fine-tuned model directly (you can route one via OpenRouter if you've hosted it there).
Mobile¶
- iOS builds are not supported (Flet target is Android-only in Whittl).
- APK builds are experimental. They work; they may need troubleshooting on specific devices or OEM skins.
- No Play Store publishing pipeline. Whittl produces signed APKs; you handle Play Console uploads manually.
- No automatic cross-platform mobile testing. Install to your own device and test there.
Code generation¶
- The AI makes mistakes. Autofix rules catch many, but not all. Some generations crash; some produce code that runs but doesn't do what you asked. Iterate.
- Large single-file projects degrade quality. Past ~1,500 lines in one file, the AI starts making more mistakes. Split into modules.
- Very domain-specific tasks (compiler writing, advanced numerical methods, hardware drivers) are usually outside the sweet spot of the base models. Whittl is best at application code.
- Real-time / low-latency requirements (audio processing under 20ms, real-time physics) are hard — Python isn't fast enough for some of these, and the AI may not pick up on your latency budget even if it understood the requirement.
Build output¶
- No code signing on Whittl's own binaries (v2.4). Windows SmartScreen warns on first install of Whittl itself; users click through "More info" → "Run anyway." Azure Trusted Signing deferred until revenue supports it. This is separate from apps YOU build with Whittl — those are also unsigned by default, and you can sign them yourself if you've got a cert.
- No automatic distribution channels. Whittl produces builds; it doesn't upload them to itch.io, Steam, the Play Store, etc. You handle distribution.
- No incremental installers. Each release is a full installer. In-app updates (v2.4+) download the full installer, not a delta.
- No macOS
.app/.dmg. Covered above.
Privacy & security edges¶
- Whittl is not encrypted at rest beyond API keys. Your code and project files are plain
.pyfiles. Full-disk encryption is the OS's job (BitLocker, LUKS). - No multi-user mode. Whittl runs per-user. On a shared machine, every user has their own
~/.whittl/and projects don't cross over. - No cloud sync built-in. Whittl is local-only. You can layer Dropbox / rclone / git on
~/.whittl/projects/to sync across machines, but nothing is built in.
Skills & Commons¶
- Skills library is text-only (markdown). No binary skill data (embeddings, vectors, code corpora).
- Whittl Commons isn't shipped yet (planned for v2.5). The "download community rules weekly" flow doesn't exist in v2.4.
What's NOT a limitation¶
To cut some common misconceptions:
- You do NOT need an internet connection if using Ollama
- You do NOT need Python installed separately
- You do NOT have to pay for the AI — Gemini and OpenRouter have free tiers usable indefinitely
- You do NOT need an Anthropic / Google / etc. account for Whittl itself — only for the backend you choose
- You do NOT need to create a Whittl account. There is no account system. The app is local.
What's next¶
- Choosing a Target — what each framework CAN do
- Choosing a Backend — backend capability comparison
- Privacy & Data Flow — explicit data flow