Follow-up: adding Ollama support to my open-source cursor-aware AI app – looking for beta testers with vision-capable local models

Follow-up: adding Ollama support to my open-source cursor-aware AI app – looking for beta testers with vision-capable local models

I’ve added support for Ollama as a first-class built-in provider in the upcoming v1.2.0 release of AIPointer. This implementation now supports:

• Auto-detection on localhost:11434
• Model dropdown populated from /api/tags
• Vision + text input pipeline (region screenshot routes to vision model)
• Tool calling for AIPointer’s 10 built-in tools (fetch_url, open_url, search_web, play_music, set_volume, copy_to_clipboard, read_clipboard, launch_app, save_document, reveal_in_finder)
• Per-model timeout (uncapped option for large models on slower hardware)
• Same config UX as the cloud providers, just point it at Ollama, pick model, done

I’ve received helpful feedback from this community regarding fast vision-capable local models. I’m now implementing support for Ollama and will need beta testers to help with testing.

What We Need From Beta Testers

• M-series Mac (M1/M2/M3/M4, Pro/Max/Ultra) – measuring TTFT against Gemini 2+3 Flash cloud baseline
• RTX 3090, 4090, or 5090 on Windows or Linux – same baseline
• AMD GPU on Linux (ROCm) – would love to know if this works at all
• 16GB-class VRAM cards – checking what’s the realistic model ceiling
• Mac mini M4 or M4 Pro – fastest consumer Apple Silicon, want to see TTFT

To participate in the beta testing, please:

1. Install AIPointer (signed + notarized on Mac, NSIS on Windows, AppImage on Linux)
2. Point it at your local Ollama, pick a vision model (Qwen2.5-VL, MiniCPM-V, Llama 3.2 Vision, Pixtral, whatever you already have running)
3. Use it for 30-60 minutes of normal daily tasks – screenshots, region queries, tool calls
4. Send back: TTFT numbers, model + quant + hardware, what worked, what didn’t, any tool-call failures

I’ll fold the feedback into the v1.2.0 release notes and credit testers/contributors if you want. If we find that one model + one inference setup consistently delivers sub-2s TTFT with reliable tool calls on consumer hardware, that becomes the recommended default in onboarding.

This is not meant to compete with any other systems; I’m building this to provide a local-inference option for people in this community. If you’re interested in participating or need more information, please let me know via DM.