Between 26 August and 7 September 2026, seven of the tools Operator Co builds on all shipped real upgrades. Here is what landed, where the source is, and whether you should update now, watch, or skip.
PRIMARY SOURCES FIRST8 RELEASES VERIFIED BY HANDNO HYPE, NO COPY-PASTE
Eight releases across seven projects. n8n also cut a 2.38.4 beta the same day, referenced in the n8n note below and in Sources, not a recommendation.
The wave landed in a tight cluster.
Eight releases from seven independent teams inside twelve days: the pattern the lead is about.
26 AugComfyUIv0.34.0
27 AugOpenHandsv1.16.0
31 AugOpen WebUIv0.11.3
31 AugCrawl4AIv0.9.3 · security
02 SepOllamav0.33.3
05 SepOllamav0.34.0-rc1 · watch
06 SepLiteLLMv1.100.0
07 Sepn8n2.37.11
Update nowWatchSecurity: update now
For a solo operator that is the whole point. The tool improves upstream, your system improves with it, and the upgrade costs nothing.
Most of this wave is unglamorous: run cleanup, migration safety, secrets that don't fall over mid-swap. That is exactly the kind of fix that decides whether an unattended automation is still running next month or died quietly at 3 a.m.
What each release actually changes for you.
Ollama02 Sep 2026
Cached-token reporting is the reliability win.
v0.33.3 · Update now
Gemma 4 takes image and audio input on the MLX engine now, Ollama reports cached prompt tokens, and it honours the default parameters baked into a GGUF. The cached-token line is the one that matters for a one-person shop: you can finally see how much of a repeat job hit the cache instead of running again: the difference between a subscription-sized bill and near-zero on your own box.
Open-weight angle: Gemma 4 is open-weight and fits on a single RTX 3090. Image and audio land on Apple Silicon first, but the model class (plus ComfyUI's Gemma 4 speedup below) is the “frontier-class, on your card” thread running through this wave.
A pre-release aimed at macOS: watch the tool handling.
v0.34.0-rc1 · Watch
This pre-release wires Ollama models straight into ChatGPT Desktop and improves structured output on Apple Silicon, plus tool search and response compaction for OpenAI-compatible clients. The headline is a macOS path, so a Windows / RTX 3090 operator gains little today. The tool-handling work is the part to watch: it carries into stable and makes your local endpoint behave more like the paid API you might be dropping.
Two reliability fixes for anything that runs unattended.
2.37.11 (stable) · Update now
A workflow run now cleans up after itself when it is rejected, and a live external-secrets provider keeps serving when its replacement fails to load. For anything that runs unattended overnight, that removes a whole category of “the job ran but wedged itself”: the failures you only find the next morning.
Pre-release 2.38.4, same day, adds the secrets fix plus a cap on task-runner timeouts and a fix for Anthropic agent threads breaking permanently. Beta only: patch what you run on stable, watch 2.38.
Frame-anchored image and audio guides for MiniMax-H3, a Gemma 4 text speedup, dynamic VRAM on by default for ROCm, HDR / AV1 / WebM video export, and Wan 3.0 video nodes. If you do client video or animation, the HDR and AV1 export plus more local video models replace a cloud GPU rental outright: same render, your card, nothing leaves the box.
A failed DB upgrade now stops instead of half-applying.
v0.11.3 · Update now
A failed database upgrade now stops at the migration error instead of half-applying and breaking later on a missing column, and chat branches survive a reload. The upgrade fix is the one that saves you a weekend: “looks fine, quietly broken” becomes a clear stop you can act on.
OpenHands now ships with an explicit skill allow-list instead of everything-on, lets you choose LLM providers from settings, and adds a Linux desktop installer and Canvas Extensions. If you let an agent touch real files, the allow-list is the change that counts: you decide exactly what it can run on the tool that edits your deliverable.
Every LiteLLM Docker image is now signed with cosign, and the release adds per-window budget spend tracking, prompt-cache savings reporting, a Bing-grounding search provider, and registry entries for new open-weight models including GLM-5.3-Flash and Kimi K2.7 Code. Signing plus cache-savings visibility is the pair that matters if you watch cost: verify the image you run, then see what caching already saved you.
Five coordinated-disclosure vulnerabilities fixed in PDF handling and the Docker playground, plus 33 accumulated bug fixes. No new features, no breaking changes. The ones that matter if you run the Crawl4AI container: a DOM-based XSS in the playground result viewer (it rendered crawled content through innerHTML), an SSRF through unchecked PDF redirects, and an arbitrary-file-write in the PDF image extractor. If your stack exposes the Crawl4AI Docker UI at all, this is not an optional bump: pull it, and check CRAWL4AI_API_TOKEN is still enforced afterwards.
€0.27per hour of real use, a used RTX 3090 amortised over a year
€7,200three years of a €200/month cloud-AI habit, owning nothing at the end
€1,850the same three years on hardware you own, all-in (card + power)
Their cloud equivalents (hosted chat, image, embedding and automation APIs) did not get cheaper. Run the same workload on a card you own and the software bill is zero. A coordinated upgrade wave like this one only widens the gap: the owned stack keeps improving at no extra cost while the subscription just renews. The full amortisation is in our €0.27/hour cloud-exit post →
Three moves, in order.
Patch the two that touch live automations first. n8n to 2.37.11 stable and Ollama to v0.33.3, restart the services, then re-run one overnight workflow so the run-cleanup and secrets fixes are on the jobs you actually depend on.
Then ComfyUI v0.34.0 and Open WebUI v0.11.3. Both are point releases, not rewrites, so low risk to pull now, and both carry a real reliability or cost win (HDR/AV1 export; a hard stop on half-finished database upgrades).
Test one cloud workload on the 3090. ComfyUI's Gemma 4 text path, or any open-weight model LiteLLM can route (GLM-5.3-Flash, Kimi K2.7 Code), and use LiteLLM's cache-savings report to write down what the local route actually cost before your vendor's next price change.
Releases collected from each project's public GitHub release feed (30-day window, zero paid fetches). Every version, date and claim was then checked by hand against the project's own release page.
One correction at that gate: the raw feed carried ComfyUI tags v0.34.1–v0.34.6. None of them exist on the canonical Comfy-Org/ComfyUI release list, which stops at v0.34.0. They are discarded. No maintainer quotes are used: every note in this wave is an auto-generated changelog, not human prose, so quoting it would add nothing.
