Ollama: MLX Runner Hardens Up, Claude Desktop Gets Real Estate

The MLX runner picked up structured output and a fix for GPU timeouts on slow storage, while a cluster of security-minded fixes closed off SSRF and CORS gaps across the app and downloader. Meanwhile, Claude Desktop integration continued its rapid rollout with icon handling, model defaults, and a toggle state fix.

Duration: PT2M42S

Episode overview

This episode is a short developer briefing from Ollama.

It explains recent repository work in plain language.

  • Show: Ollama
  • Published: 2026-08-26T13:01:12Z
  • Audio duration: PT2M42S

Transcript excerpt

This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.

It's August 26th, and this is your Ollama briefing.

The clearest signal today: MLX is catching up to feature parity with the llama-server path, and several PRs closed real security gaps around network boundaries.

Start with MLX. Jesse Gross landed structured output support in PR 17929 — the runner now actually enforces the format field using a grammar engine, so JSON and JSON Schema requests get valid output instead of silently unconstrained text. Companion PR 17998 fixes a nastier problem: on Metal, loading large models…

Second theme: security boundaries. PR 18021 blocks the downloader from following redirects to private or link-local addresses — closing an SSRF path where a malicious registry could redirect a tensor download to something like a cloud metadata endpoint. PR 17989 rejects browser origins on the Claude Desktop gateway…

Third theme: Claude Desktop keeps expanding. Model default resets, icon handling that pulls from the installed app instead of bundling one, a toggle that no longer flickers mid-connection, and a proxy fix so string-based message content survives the image fallback path. Docs landed same-day in PR 18004.

Smaller but worth noting: PR 18017 trims whitespace…

Nearby episodes from Ollama

  1. Claude Desktop Integration Overhaul
  2. Hardening the Runner Scheduler
  3. Weekly Recap - Desktop Apps Grow Up, Inference Gets More Reliable
  4. Visibility Into What the Model Actually Sees
  5. Fixing the Long-Prompt Timeout Trap
  6. Claude Desktop Gets Deep App Integration
  7. Closing the Gaps on Hung Requests and Broken Installs
  8. Closing the Gap on Model Metadata Overhead