Dyno Lab 0.6.18

Dyno Lab 0.6.18 puts a model you serve online, still on your Mac: with Web access it can search the web and read pages, and the new Browser tab lets it click and type, as far as you allow. GGUF files now run on this Mac too, with llama.cpp on the GPU, in Models, dyno serve and dyno launch.

Browser

• A browser the model can use. The new Browser tab is a web browser that models with Web access drive, like Opera's browser agent: they open pages, read them as text and numbered links, buttons and fields, and, if you allow it, click and type. Choose Read only (the default), Ask before submitting (Dyno asks you before anything that submits, buys, sends, posts or deletes) or Full control. It never types passwords or card details, keeps its own cookies and logins, opens public pages only, and lists every step; Pause the model lets you take over.

Models

• GGUF models on this Mac. Models → GGUF now lists your downloaded GGUF files with the same Start panel as MLX models. Dyno runs them with the llama.cpp build in the app, on the GPU, behind its own endpoint, so Chat, Claude Code and Codex (dyno launch), Web access, the Browser, the Observer and Performance work with them too. From the terminal: dyno serve --model file.gguf. Before, GGUF files could only run in Pools.
• Web access. A new Web access switch beside Start (dyno serve --web) lets apps that chat with a served model, such as chat apps that bring no tools of their own, search the web and read pages. Dyno gives the model web_search and read_page, runs them on this Mac with the search from Settings → Web search, and the app gets the finished answer. Pages on this Mac and your local network can't be read. Each search and page read shows in Performance → Logs. Off by default. Claude's web fetch tool is now answered the same way as its web search.

Command line and MCP

• GGUF models in dyno models and dyno launch. dyno models lists the GGUF files you downloaded in Models → GGUF, and GGUF repositories in the Hugging Face cache, so dyno launch claude --model qwen3-0.6b-q8 runs Claude Code or Codex on one. dyno serve --model file.gguf no longer needs MLX installed.
• Why a server can't capture activations. /lab/capabilities, the SDK's ServingModel and the MCP tools serving_capabilities and serving_inspect now say why capture isn't available: the model is a GGUF file served by llama.cpp, or the server runs in serve-only mode.