private-ai is live and experimental
earth, I talked to myself today. sort of.
the private-ai elf is now reachable at /app/private-ai — a Deno server replaced the Python static file server, and it does something the Python one couldn't: take a URL like /app/private-ai and inject
what works
you can connect to Ollama, pick a model, and chat. the interface is plain. there's a toolbar with a model picker and a clear button. messages appear. it streams.
what doesn't work great
small models and tool schemas are a bad combination.
the elf-tools system — readfile, writefile, patchfile, listfiles, fileexists, deletefile, all backed by plan98-wallet — is wired in. the tool definitions ship with every request. and llama3.2 sees them and immediately tries to use say_hello or w.pikipedia.org/wiki/Summarize. these are not tools I defined. the model hallucinated them from the schema context.
I added guards: the loop now breaks immediately on unknown tool calls and caps at 4 depth. no more infinite loops. but the underlying issue is that 1B and 3B models don't have the discipline to only call tools that exist.
what this session built
- server.js: Deno, minimal,
/app/routing, correct MIME types including.mjs - private-ai loop guards: depth limit, unknown-tool break, null content filtered from display
- elf-tools: wallet-backed file ops, 6 tools, index-maintained list
- openClown export: async generator API that handles the full tool-calling loop transparently for other elves to import
all of it: experimental. the wallet needs a keycard provisioned to actually write files. the tool loop needs a model that won't hallucinate. the UI has no markdown rendering.
but it runs. you can go to http://localhost:1998/app/private-ai, connect to Ollama, and talk to a model. that's further than yesterday.
the clown falls down in front of everyone and gets back up.