https://toolsift.ai/new/ Short, hand-written ToolSift verdicts on notable GitHub releases of AI, LLM and agent tools: what changed and whether to update. Editorial, never paid for. Updated 2026-10-21.en-gbWed, 21 Oct 2026 07:59:04 GMT https://toolsift.ai/new/#n8n-n8n-2-44-5toolsift-take:n8n-io/n8n@n8n@2.44.5Wed, 21 Oct 2026 07:26:46 GMTn8n-io/n8nToolSift take: n8n's stable tag now points at 2.44.5, so 2.44 replaces 2.43 as the stable line. The patch itself has two fixes: queue job outcomes are keyed by job when resuming after a shutdown, and the editor loads credential types on startup. Source: https://github.com/n8n-io/n8n/releases/tag/n8n%402.44.5 https://toolsift.ai/new/#claude-code-v2-1-306toolsift-take:anthropics/claude-code@v2.1.306Wed, 21 Oct 2026 03:16:31 GMTanthropics/claude-codeToolSift take: A two-fix patch: claude -p inside cloud sessions no longer reports "Not logged in" when started without the session's environment variables, and sessions with many MCP tools no longer have an API error replaced by "prompt is too long" plus an unneeded compaction. Source: https://github.com/anthropics/claude-code/releases/tag/v2.1.306 https://toolsift.ai/new/#openhands-v1-29-0toolsift-take:OpenHands/OpenHands@v1.29.0Tue, 20 Oct 2026 22:22:52 GMTOpenHands/OpenHandsToolSift take: The theme is agent profiles: each one can now choose its own skills from the server catalog, Model Router, condenser and verification settings. Fixes include the chat catching up with the model within a second instead of ten and keeping the first Cloud message when the history page lags; the macOS app is still unsigned, so Gatekeeper may call it "damaged" until you clear the quarantine flag as the notes describe. Source: https://github.com/OpenHands/OpenHands/releases/tag/v1.29.0 https://toolsift.ai/new/#gemini-cli-v0-65-0toolsift-take:google-gemini/gemini-cli@v0.65.0Tue, 20 Oct 2026 21:17:57 GMTgoogle-gemini/gemini-cliToolSift take: Almost all fixes, several worth having: read-only workspace settings are now enforced in untrusted folders, OAuth callback iss validation follows RFC 9207, and auth no longer gets stuck in infinite verification or OAuth retry loops. Resumed sessions also stop duplicating tool responses or losing history on a quick exit; the one feature is custom OTLP headers for telemetry. Source: https://github.com/google-gemini/gemini-cli/releases/tag/v0.65.0 https://toolsift.ai/new/#claude-code-v2-1-305toolsift-take:anthropics/claude-code@v2.1.305Tue, 20 Oct 2026 20:46:50 GMTanthropics/claude-codeToolSift take: Update if you're on 2.1.304 with Vertex, Bedrock, another provider or a gateway: every request could fail when the endpoint or gateway rejected tool search, and Claude Code now retries without it. The new sandbox.allowNestedUserNamespaces setting lets Linux sandboxed commands such as Chromium or rootless podman create user namespaces, but the notes say this lets them get around the sandbox's file protections, so leave it off unless you need it. Source: https://github.com/anthropics/claude-code/releases/tag/v2.1.305 https://toolsift.ai/new/#n8n-n8n-2-45-0toolsift-take:n8n-io/n8n@n8n@2.45.0Tue, 20 Oct 2026 09:10:14 GMTn8n-io/n8nToolSift take: A large beta pre-release with default changes to note before it reaches stable: the durable scheduler and node and credential type policies are enabled by default, removed nodes get automated migrations, and MCP connections are no longer saved without user approval. Public API POST requests also honour Idempotency-Key; test it before moving production onto it. Source: https://github.com/n8n-io/n8n/releases/tag/n8n%402.45.0 https://toolsift.ai/new/#claude-code-v2-1-304toolsift-take:anthropics/claude-code@v2.1.304Mon, 19 Oct 2026 21:34:39 GMTanthropics/claude-codeToolSift take: A big fix release: update if you use worktrees, because claude --worktree could remove the worktree and its branch without asking at exit when HEAD was detached or on another branch, which could lose commits. New in this one: a guided /auto-mode-setup (now also on Bedrock, Vertex and Foundry) and Quick, Standard and Deep levels for Code Review. Source: https://github.com/anthropics/claude-code/releases/tag/v2.1.304 https://toolsift.ai/new/#transformers-v5-20-2toolsift-take:huggingface/transformers@v5.20.2Mon, 19 Oct 2026 17:45:37 GMThuggingface/transformersToolSift take: Worth taking if you run multimodal models: some inputs could lose their original type inside prepare_inputs_for_generation during generate. It also fixes zero expert gradients in bfloat16 with expert parallelism when ep_size is smaller than fsdp_size. Source: https://github.com/huggingface/transformers/releases/tag/v5.20.2 https://toolsift.ai/new/#openhands-v1-28-0toolsift-take:OpenHands/OpenHands@v1.28.0Mon, 19 Oct 2026 14:29:10 GMTOpenHands/OpenHandsToolSift take: Mostly polish, with dozens of theme, accessibility and phone-layout fixes; the headline feature is that the agent can now ask for a secret inline in the conversation. Cloud users also get fixes for messages failing before the runtime is reachable and for resuming a paused sandbox, and 1.28.1 followed with an agent-server bump to 1.56.0. Source: https://github.com/OpenHands/OpenHands/releases/tag/v1.28.0 https://toolsift.ai/new/#langgraph-cli-0-4-37toolsift-take:langchain-ai/langgraph@cli==0.4.37Mon, 19 Oct 2026 14:06:22 GMTlangchain-ai/langgraphToolSift take: A small CLI release with one change to check: deploy now defaults to serverless, so confirm your deploy target if you script deployments. Verbose output also no longer prints subprocess stdin. Source: https://github.com/langchain-ai/langgraph/releases/tag/cli%3D%3D0.4.37 https://toolsift.ai/new/#servers-v2-0-0toolsift-take:modelcontextprotocol/servers@v2.0.0Mon, 19 Oct 2026 06:46:48 GMTmodelcontextprotocol/serversToolSift take: A big one if you run the official MCP servers: everything, filesystem, memory and sequential-thinking are rebuilt on the version 2 TypeScript SDK for the modern-era spec, while still answering older clients. Check two things before upgrading: the TypeScript servers now need Node 24 or later, and the Python servers (fetch, git, time) now return a JSON-RPC error for an unknown tool name instead of an error result. Source: https://github.com/modelcontextprotocol/servers/releases/tag/v2.0.0 https://toolsift.ai/new/#comfyui-v0-42-1toolsift-take:Comfy-Org/ComfyUI@v0.42.1Mon, 19 Oct 2026 01:33:54 GMTComfy-Org/ComfyUIToolSift take: A one-fix patch for the crashes some users hit running the MOSS image template workflow; update if you're on 0.42.0 and use MOSS-image. Source: https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.42.1 https://toolsift.ai/new/#comfyui-v0-42-0toolsift-take:Comfy-Org/ComfyUI@v0.42.0Sun, 18 Oct 2026 22:10:38 GMTComfy-Org/ComfyUIToolSift take: Adds MOSS-image model support and a Krea 3 Image partner node, speeds up Qwen 3.x, and fixes the webcam node and a solid-mode crash. Some users hit crashes with the MOSS image template, so go straight to 0.42.1, which fixes them. Source: https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.42.0 https://toolsift.ai/new/#vllm-v0-32-0toolsift-take:vllm-project/vllm@v0.32.0Sun, 18 Oct 2026 06:50:48 GMTvllm-project/vllmToolSift take: Worth it for serious self-hosters: low-precision KV cache on more GPUs, faster startup with precompiled FlashInfer kernels, new models such as EmbeddingGemma 2 and MiniCPM-V 4.7, and three security advisories fixed. Read the breaking changes first: Python 3.10 is dropped, CORS no longer allows every origin by default, and max_tokens is now only an upper bound on output. Source: https://github.com/vllm-project/vllm/releases/tag/v0.32.0 https://toolsift.ai/new/#cline-desktop-v0-0-48toolsift-take:cline/cline@desktop-v0.0.48Sat, 17 Oct 2026 04:34:20 GMTcline/clineToolSift take: If you use GPT-6 models through OpenAI or Azure OpenAI-compatible endpoints, update: every request was being rejected because Cline sent the wrong token-limit field. It also makes Claude Sonnet 5.5 the Cline provider's default again and refreshes the model catalogue. Source: https://github.com/cline/cline/releases/tag/desktop-v0.0.48 https://toolsift.ai/new/#langgraph-1-2-16toolsift-take:langchain-ai/langgraph@1.2.16Fri, 16 Oct 2026 22:11:28 GMTlangchain-ai/langgraphToolSift take: A fix release that matters if you use interrupts and resume: resumed commands now apply their update once, finished subgraph tasks no longer rerun, and an invalid resume value no longer stops other tasks. The Postgres checkpointer also gains a prune method for old checkpoints. Source: https://github.com/langchain-ai/langgraph/releases/tag/1.2.16 https://toolsift.ai/new/#ollama-v0-40-6toolsift-take:ollama/ollama@v0.40.6Fri, 16 Oct 2026 22:00:46 GMTollama/ollamaToolSift take: A light update: it adds the q2_0 GGUF tensor and file types and stops runner processes being left orphaned on Windows. Update if you're on Windows; otherwise there's no rush. Source: https://github.com/ollama/ollama/releases/tag/v0.40.6 https://toolsift.ai/new/#open-webui-v0-12-3toolsift-take:open-webui/open-webui@v0.12.3Fri, 16 Oct 2026 11:46:21 GMTopen-webui/open-webuiToolSift take: A meaty point release: admins can switch on temporary end-to-end encrypted chats in DMs and private channels (off by default), streaming with Redis is much lighter, and skills attached to a model now send their full instructions. It also fixes OAuth token refresh for providers that only accept client_secret_basic, which was silently breaking connections. Source: https://github.com/open-webui/open-webui/releases/tag/v0.12.3 https://toolsift.ai/new/#opencode-v1-19-2toolsift-take:anomalyco/opencode@v1.19.2Fri, 16 Oct 2026 02:30:38 GMTanomalyco/opencodeToolSift take: If you're on OpenCode 1, update checks now only offer 1.x builds, and moving to OpenCode 2 is a separate, explicit step through the 2.0 offer in /update or opencode upgrade --major. Homebrew Core installs now track OpenCode 2 and are skipped for 1.x updates. Source: https://github.com/anomalyco/opencode/releases/tag/v1.19.2 https://toolsift.ai/new/#browser-use-0-13-13toolsift-take:browser-use/browser-use@0.13.13Thu, 15 Oct 2026 23:50:45 GMTbrowser-use/browser-useToolSift take: Adds an OpenAI browser-tools integration that gives OpenAI's Python SDK 18 default browser tools on a local browser or Browser Use Cloud. It doesn't do uploads, downloads or Bash yet. Install 0.13.13 rather than 0.13.12, whose PyPI release was cancelled. Source: https://github.com/browser-use/browser-use/releases/tag/0.13.13 https://toolsift.ai/new/#comfyui-v0-41-0toolsift-take:Comfy-Org/ComfyUI@v0.41.0Thu, 15 Oct 2026 22:57:45 GMTComfy-Org/ComfyUIToolSift take: Adds Flux3-image model support, an SDR-to-HDR image node, a ReferenceMultiLatent node and MiniMax-H3 LightVAE support, plus updated workflow templates. Worth updating if you want to try Flux3; otherwise routine. Source: https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.41.0 https://toolsift.ai/new/#codex-rust-v0-163-0toolsift-take:openai/codex@rust-v0.163.0Thu, 15 Oct 2026 21:30:38 GMTopenai/codexToolSift take: A solid quality-of-life release: send a follow-up while Codex is responding to steer it straight away, search settings from a new /config screen, and independent read-only lookups now run in parallel. It also tightens the network proxy (permission checks before DNS lookups, duplicate Host headers rejected), so it's worth updating. Source: https://github.com/openai/codex/releases/tag/rust-v0.163.0 https://toolsift.ai/new/#opencode-v1-19-1toolsift-take:anomalyco/opencode@v1.19.1Thu, 15 Oct 2026 19:19:45 GMTanomalyco/opencodeToolSift take: A heads-up if you share sessions: session sharing is discontinued, so existing share links and session-sharing commands no longer work, and importing sessions from share URLs is unsupported. Source: https://github.com/anomalyco/opencode/releases/tag/v1.19.1 https://toolsift.ai/new/#transformers-v5-20-0toolsift-take:huggingface/transformers@v5.20.0Thu, 15 Oct 2026 18:05:36 GMThuggingface/transformersToolSift take: Adds support for Meta's Muse Spark Open (816B total, 42B active parameters, up to 1M-token context) plus six other new models, including a 3B OCR model from Baidu and a speech transcription-and-diarisation model. It has breaking changes and the notes warn of a speed regression in expert-parallel inference, so pin your version and test before upgrading production code. Source: https://github.com/huggingface/transformers/releases/tag/v5.20.0 https://toolsift.ai/new/#ollama-v0-40-5toolsift-take:ollama/ollama@v0.40.5Wed, 14 Oct 2026 19:45:22 GMTollama/ollamaToolSift take: Adds NVIDIA's Nemotron 3.5 Super, a 120B mixture-of-experts vision-language model (12B active) for coding and tool use, plus a smaller Nemotron 3.5 Bolt sibling that NVIDIA's notes say is tuned to fit a 24 GB GPU. Decision models can now run on Ollama's cloud as well as locally. Update if you want the new models. Source: https://github.com/ollama/ollama/releases/tag/v0.40.5 https://toolsift.ai/new/#servers-v1-0-0toolsift-take:modelcontextprotocol/servers@v1.0.0Sun, 11 Oct 2026 02:40:29 GMTmodelcontextprotocol/serversToolSift take: The official reference MCP servers were republished: the npm ones (filesystem, memory, sequential thinking and the 'everything' test server) reach 1.0, while the Python ones (fetch, git, time) are date-versioned 2026.10.10. A good stability marker if you wire tools into MCP clients. They are reference servers, so check what folders and permissions you give the filesystem one. Source: https://github.com/modelcontextprotocol/servers/releases/tag/v1.0.0 https://toolsift.ai/new/#google-research-harness2toolsift-take:google-research/harness2Tue, 06 Oct 2026 21:48:11 GMTgoogle-research/harness2ToolSift take: Research code, not a product: Google Cloud AI Research and Northwestern's Harness² (arXiv:2610.14224, 9 Oct 2026) improves an agent's surrounding harness (prompts, tools, memory, skills) rather than the model, by proposing edits, probing them on the same task and keeping what helps, then carrying lessons to later tasks. The paper reports gains of up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench and 11.8 percentage points on LAB for GLM-5 and Gemini 3.5 Flash. Those are the authors' own numbers; the Apache-2.0 release runs with OpenCode or Codex executors. Source: https://github.com/google-research/harness2 Paper: arXiv:2610.14224 (https://arxiv.org/abs/2610.14224)