SynthPanel – a test agent community for everyone building with AI

A lot of people are building products and services with AI right now. The problem is always the same: who tests it? Real users are expensive. Your own perspective has blind spots. And a quick „looks good“ from a friend isn’t user research.

SynthPanel solves that. It’s open source, it runs locally, and it’s on GitHub today.

What it does

You give it a short description of your target audience and the thing you want tested – a pitch, a UI, a speech, an app idea, a piece of text. SynthPanel generates a large swarm of diverse test personas and runs them against your artifact in parallel.

Each persona comes from a different angle: different demographics, different technical backgrounds, different assumptions. The results get synthesized into structured feedback – not a wall of opinions, but a ranked view of where the blind spots are and what actually lands.

How it works

There are two layers. A main orchestrator agent runs the intake conversation with you, builds the persona set, and synthesizes the results. A cheap parallel swarm does the actual testing – each persona runs independently, which keeps costs down and coverage high.

Locally it runs on Ollama. We recommend qwen3.5 – good balance of speed and reasoning depth. llama3.2:3b works for quick runs. If you don’t have a local setup, it falls back to Claude agents – but that will burn tokens fast, so consider yourself warned.

There’s also a GitHub Action included. Drop it into any repo and it runs the swarm as part of your CI – useful for catching regressions in UX copy, accessibility, or tone before they ship.

Optional: accessibility

One of the persona dimensions is accessibility. Screen reader users, color blindness, motor impairments – the swarm can cover those perspectives explicitly if you ask for it. Not as a checkbox, but as actual test feedback from a persona that has those constraints.

Where to find it

GitHub: github.com/spm231177-coder/synthpanel

MIT license. Works with local Ollama or Claude. Pull requests welcome.

If you try it: what did the swarm catch that you missed? Let me know.