LokalBot update: local Agent Mode, MCP access to your meetings, system-wide dictation and notarized build, still 100% on-device & open source

Two weeks ago I posted LokalBot here (previous thread) the Mac app that records + summarizes meetings, autocompletes your typing in any app, and tracks your day, with every model running on-device. The feedback in that thread turned into a two-week fix-and-ship sprint, so here's an update. First v0.2.0 release is properly signed as I purchased the Apple Developer account, no more Gatekeeper warnings. Second, the RAM spikes I warned about last time. All built-in inference now goes through a broker: the three llama-server instances (main / embeddings / cotyping) are lease-managed, so anything with no active work unloads after a short linger and gives the RAM back, and loaded models live under a residency budget with LRU eviction. There's a live resource monitor in Settings showing exactly what's loaded and which lease is pinning it. Idle footprint is no longer "whatever you touched last" servers only hold memory while something is actually using them. New since the last post: Agent Mode. An embedded coding agent (pi) wired to the same local Main LLM through an OpenAI-compatible shim i.e. a coding agent running on your own GGUF, fully offline (--offline, version checks disabled). write / edit / bash calls are gated behind native approval cards. The Bun runtime it needs is checksum-pinned and downloaded on first enable, not bundled. MCP server for your meeting library. lokalbot-cli mcp serves list_meetings / get_meeting / search_meetings / ask_library over stdio, so Claude Desktop (one-click .mcpb bundle) or any MCP client can query your meetings. ask_library answers through the app's local llama-server so even "chat with my meetings from another app" never leaves the machine. Off by default, behind an explicit Privacy toggle. System-wide dictation. Hold ⌥ Space, talk, release on-device ASR pastes at the cursor in any app. Pauses your music first, shows a live transcript while you speak, deletes the audio after. Reuses whichever transcription engine + language you already picked. Live meeting view. Rolling transcript + quick notes during the call (opt-in), instead of everything arriving only after it ends. On-device TTS. Kokoro (via sherpa-onnx) reads summaries and chat answers aloud. Cotyping fixes: mid-word completion, caret finally renders correctly in Chromium/Electron apps, better sentence/punctuation handling. Boring but important: model downloads are now resumable + SHA-verified + disk/RAM-prechecked; recording recovery got hardened so a crash mid-meeting shouldn't cost you the audio; chat tool calls are JSON-schema-constrained and answers cite the source meeting; the remote speaker gets named from the calendar invite instead of "Them." UI reorg: sidebar collapsed to Capture / Ask / Type, onboarding only requires the mic, and screen capture / day tracking are true opt-ins now. Unchanged: Apple Silicon / macOS 15+ only (it's welded to the Neural Engine, MLX, and Core Audio process taps no Linux/NVIDIA, sorry). Defaults are still Granite Speech 4.1 (2B) Q4 for transcription, Qwen 3.6 35B-A3B for summaries, Gemma 4 E4B for cotyping nobody in the last thread talked me out of them, so here's your second chance. Precision on the network claim, since this sub keeps me honest: out of the box the only traffic is the one-time model/runtime downloads (Hugging Face for GGUFs, GitHub for the pinned llama.cpp / sherpa-onnx binaries, plus Bun if you enable Agent Mode). After that it's fully offline and every server is localhost-only. Little Snitch during a meeting still shows the flattest network graph you've ever seen. GitHub (GPLv3): https://github.com/stevyhacker/lokalbot Landing: https://lokalbot.com My main asks: a better small cotyping model fast enough for keystroke-latency ghost text, smart enough to be worth pressing Tab for. And if you point your own agent setup at the MCP server, I'd love to hear what tools are missing.