FLVSTX 1.3.0 — Ollama

Compose on an Ollama model — the one on this PC or one across the room — instead of Claude.

Ollama

Pick Ollama beside Claude in the chat header, type the server URL (http://localhost:11434
is the Ollama on this PC; another machine works if Ollama there runs with OLLAMA_HOST=0.0.0.0),
and choose a model from the list the server reports. ↻ re-asks after an ollama pull. The
choice is saved with the project, and Claude keeps its own conversation, so switching back resumes
where it left off. No cost is shown for local turns: nothing leaves your machine.

Think lets a thinking model reason before it answers. Off by default — on a small GPU it is
minutes of deliberation before the first tool call — and when it is on, the reasoning streams into
the transcript and collapses once the answer arrives. Claude's thinking summaries show the same way.

Terminal chat: /ollama [URL], /models, /model NAME, /think on|off, /claude.

What to expect: pick a model with tool calling (ollama show NAME lists tools); every turn carries
about 10k tokens of brief and tool schemas, so very small models lose the thread and a model that
does not fit your GPU crawls on the CPU. On an 8 GB card a 27B model is a multi-minute turn; an 8B
one (ollama pull qwen3:8b) is well under a minute. Producer mode runs on Ollama too, with the same
checklist catching a specialist that skips a step.

Under the hood the sidecar hosts a small bridge onto Ollama's native chat API rather than its
Anthropic-compatible endpoint, which cannot set the context window and truncated the composer's
brief to 2048 tokens. FLVSTX_OLLAMA_NUM_CTX (default 16384, capped at the model's limit),
FLVSTX_OLLAMA_URL, FLVSTX_PROVIDER=ollama and FLVSTX_OLLAMA_THINK=1 set the sidecar's defaults.

Install

Unzip, then run scripts\install.ps1. Node 22 is needed for the chat composer; Claude Code login
for Claude, or a running Ollama for local models.