Ollama v0.33.0-rc2 wires local models into Claude Desktop and fixes a KV-cache break caused by Claude Code's own token countdown

GitHub·high signalXBlueskyLinkedInCopy link

The release candidate published 2026-08-21 adds a Claude integration: individual Ollama models can be toggled on or off for use inside Claude from the menu bar, with cloud models appearing only when signed in, plus a new Apps view for managing integrations. The caching section is the more interesting part for builders — Ollama disabled Claude Code's 'tokens left' countdown system message because Ollama was moving it to the front of the prompt and invalidating the KV cache on every single request. It also fixed cancelled prefills discarding restore points, a bug that on recurrent-layer models forced a request matching 46k of 47k tokens to reprocess from zero.