Software Engineer & AI Enthusiast
Polyphon is now open source under Apache 2.0. The reason isn't contributor growth or community optics — it's that local-first AI software can't ask users to accept privacy and control claims on faith alone. Open source gives users two things: the ability to inspect the architecture and verify what the tool is actually doing, and a real exit if the project ever changes direction.
When I shipped Polyphon v0.1.0-alpha.2, session export and a plugin system were next on the roadmap. Today it's v0.8.0, and the features that actually changed how the product feels weren't the ones I planned: voices that interact with real files with per-voice sandboxing, whole-database AES-256 encryption with password protection, and FTS5 search that turns conversation history into working memory. This is the release where Polyphon starts to feel trustworthy for real work.
Polyphon just shipped MCP server support, and it changes the product's role in a developer workflow. Instead of opening Polyphon as a UI you visit deliberately, your coding agent — Claude Code, Cursor, Codex CLI, GitHub Copilot, Windsurf, Gemini CLI — can now call into it directly and broadcast a prompt across your entire Claude + GPT-4o + Gemini ensemble in one shot. Here's what shifted, why the SessionEventSink abstraction made it possible, and when multi-model review is actually worth it.
For decades, 'free software' and 'open source' were nearly synonymous — because building something worth sharing required a community. AI has dissolved that constraint. But the story isn't that open source is dying. It's that open source used to do five jobs at once, and AI is separating those threads. Some become less necessary. Others become more critical than ever.
Polyphon is a desktop app I shipped this week for orchestrating multiple AI models in a single conversation. Instead of tab-juggling between Claude, GPT, and Gemini, you run all of them as voices in one shared session where they can actually read and respond to each other. Here's why I built it, what technical bets I made, and what I learned on launch day.
This blog runs on a custom automation pipeline where AI agents compete, critique, and synthesize every post. Here's how the /create-blog-post command, a DALL-E 3 image generation script, pre-commit image enforcement, and WebP optimization all compose into a content system that behaves like production software.
Read more
AI agents can write code, fix bugs, and ship features — but who can be trusted to ship production software when most implementation is delegated? Meet the Agent Wrangler: the engineer who decomposes work, directs multiple agents, validates output, and owns the outcome. Deep technical foundations matter more than ever. The keyboard gets less important. Judgment gets everything.
After refusing to pay for yet another AI subscription I couldn't customize, I switched from Notion to Obsidian and discovered there was no way to use my existing AI services inside it. So I built one — an open source AI Agent Sidebar plugin developed with agentic engineering and the dark factory method.
The first version of agent-config tried to impose uniform rules on Claude, Codex, Copilot, and Gemini — and it broke. The fix wasn't better engineering up front. It was accepting failure fast, letting AI handle per-agent format conversion, and shipping a v2 that actually works. Here's what the failure taught about human instinct, AI-assisted iteration, and why failing cheaper is a skill worth building.
Between Claude Pro, ChatGPT Plus, GitHub Copilot, Cursor, and a handful of specialty tools, a serious AI stack can run $100 or more per month. This post explores a practical closed-loop strategy: give every subscription a job, pick one small revenue experiment, and use the tools themselves to generate enough income to cover the bill.