Inspiration

Every productivity app makes you fit its idea of how you should work, and most AI tools tie you to one company's model in the cloud. We wanted the opposite: describe the tool you need, have it built in minutes, and stay in control of which AI builds it. You pick from any model on your own computer, so you can choose speed, quality or privacy without changing a thing.

What it does

WorkBench is an AI tool builder that works with whatever AI model you want. Any model you run through Ollama, such as Qwen, Llama, DeepSeek or Mistral, appears in a menu at the top of the page. Switch between them at any time:

  • Fast: a small model (e.g. 7B) for simple tools in seconds.
  • Powerful: a large coding model (e.g. qwen2.5-coder:32b) for complex tools like calendars and kanban boards.
  • Your own: any fine-tuned or custom model you have installed.

You describe a tool in plain English, and the model you picked:

  1. Plans it: writes a short spec, so you can see what it's about to build.
  2. Builds it: writes a complete, self-contained tool, and you watch the code appear live.
  3. Tests it: loads the tool in a hidden sandbox, clicks every button and catches errors.
  4. Fixes it: if anything breaks, sends the errors back to the AI to fix automatically.

If you like the result, Keep it saves it to your library along with its data, so a calendar remembers its events. You can then change it in plain words ("add a dark mode"), restart it, clear its data or download it as its own HTML file. Everything can run fully offline, so your work never leaves your machine.

How we built it

  • Model-agnostic backend: Flask (Python) talks to Ollama, so any model it can run works with WorkBench, with no code changes. The server lists the installed models and suggests the best one for coding. Your choice is remembered.
  • Streaming: replies stream back as they're written, so you see progress with any model, fast or slow.
  • Frontend: one HTML file of plain JavaScript and CSS, with no frameworks.
  • Sandbox: each generated tool runs in a locked-down iframe. A small injected script gives the tool working storage that WorkBench saves for it, reports errors back, and runs the automatic click test.
  • Model-independent prompting: a planner prompt and then a strict builder prompt (one file, no internet, save data, no placeholders). Together with the test-and-fix loop, this lets even smaller models produce working tools. -**We also employed AI

Challenges we ran into

  • Models vary a lot. A general model like llama3 handled a stopwatch but broke on a calendar. Rather than lock users into one model, we made the pipeline itself more reliable: planning, then automatic testing and fixing, so every model does better.
  • Big tools got cut off when the code outgrew how much text a model can handle at once. We now give models $$4\times$$ more room and detect when code ends early.
  • Speed versus quality. Large models are slow on a laptop: a full calendar can take 10+ minutes on a 32B model. Letting users switch models turned this from a limitation into a choice.
  • Browser security. Ollama blocks pages opened straight from a file, and on macOS port 5000 is taken by AirPlay. The page now detects these and tells you the exact fix.
  • Saving data from a sandbox. Sandboxed tools can't use the browser's storage, so we built a bridge that passes their data to WorkBench to save.

Accomplishments that we're proud of

  • Bring your own AI: no lock-in to one provider or model. Pick it from a menu.
  • Runs fully offline: no API keys, no cloud, no data leaving your machine.
  • Tools test and repair themselves before you ever see them, whichever model built them.
  • Kept tools remember their data, so they're real tools and not throwaway demos.

What we learned

  • Different models suit different jobs, so users should get to choose instead of being stuck with one.
  • A coding model beats a general one at the same size.
  • Self-testing works with any model: running the code and sending real errors back fixes more bugs than asking the model to "be careful".
  • Clear error messages matter as much as features.

What's next for WorkBench

  • Even more AI choices: support for OpenAI-compatible endpoints (LM Studio, llama.cpp servers and optional cloud APIs), so you can use literally any model you want.
  • Per-task model mixing: a fast model for planning and a strong one for code, chosen automatically or by you.
  • Faster fixes: have the AI send back only the changed lines instead of rewriting the whole file.
  • Accounts and syncing, so your library follows you across devices.
  • Tools working together, for example a timer that logs into a habit tracker, plus a starter gallery of ready-made tools.

Built With

Share this project:

Updates

Submission history