Skip to main content

Explaining Synfire to an AI

Thu, 2026-11-12 - 08:55 Permalink

Of course. I'm now testing compatibility with different clients. Claude Desktop works best. OpenAI's desktop client (via Codex MCP) also works fine, although the app is really creepy (installs a lot of unwanted software that takes over your computer). 

Even local AI does fine:

An old Mac M1 with 32GB can run this Qwen model locally. It understands the Synfire tools, properly uses them, and does surprisingly fine. But it's not nearly as clever as the frontiers. I have to babysit it and explain things that the other models figure out on their own.

Dumber models are a good test for optimizing the tool descriptions and instructions though.

Tue, 2026-11-17 - 19:15 Permalink

  "MCP tools are now complete. Countdown is running ..." 
lol, still counting 

Wed, 2026-11-18 - 11:02 Permalink

Workload is currently brutal. Testing doesn't mean it all works flawlessly aready. That last 1% is the hardest. Some LLM clients are still acting up. The connector won't probably work with every local client. We tested Claude Desktop, the ChatGPT app, llama.cpp local servers, and MCP inspector with protocol versions going back to mid 2025. Is anyone using Google's Gemini?

I did a video yesterday. It's 40 minutes – no idea yet how to compact that to make it watchable. The result was fantastic but agentic workflows run longer than anyone's attention span. Time lapse misses important things, like how it uses all the different Synfire features. Much like with Synfire itself, this is an overwhelming amount of new information to explain and present.

With all the instructions/briefings markdown docs, persistent agent memory inside Synfire, the SQL database for phrase similarity matching, etc this turns out to be as challenging as a new product.

I'll start a thread covering this shortly.