Ollama v0.34.0-rc1 Brings Ollama Models to ChatGPT Desktop
What shipped
Ollama v0.34.0-rc1, published on GitHub at 23:49 UTC on September 5, 2026 — the first release candidate of the 0.34 line, ten commits ahead of stable v0.33.3. Headline: Ollama models can now be used directly in ChatGPT Desktop, with setup from the Ollama app on macOS.
What changed
- The integration itself: PR #18236 ("app: add Ollama to ChatGPT Desktop") merged September 4 at 21:04 UTC — 11 days after Ollama shipped the identical deal for Claude Desktop on August 25.
- API plumbing the desktop client needs: Ollama's local server now speaks OpenAI-compatible client tool search (PR #18223) and response compaction (PR #18224), with images preserved through compacted responses (PR #18245).
- Desktop hardening: proxy handling for the Codex desktop path (PR #18244) and auto-review routed through the selected model (PR #18246).
- Also: structured-output performance improvements on Apple Silicon.
Why a builder cares
The number that matters is 11 days. Anthropic opened the door on August 25; OpenAI walked through it inside two weeks. The frontier desktop apps are converging on the same architecture — a shell over whatever local runtime you point at them. That is a distribution win for Ollama (180,272 GitHub stars, 239 model pages in its library) and slow moat erosion for LM Studio-class GUIs whose whole pitch was "frontier app plus local models."
Second unlock, no GUI required: agent frameworks built on OpenAI SDKs now run against localhost with zero code changes, because tool search and compaction — the two Responses-API features agents lean on hardest — are implemented. One caution before you trust it: PR #18244 exists because proxies sit inside this loop. Check what leaves your machine.