LokalBot 0.10.8

What’s new

• Close the window and keep working. Closing LokalBot’s main window now leaves the app running, so background capture can continue. Reopen it from the menu bar or Dock; use Quit or ⌘Q when you want to stop the app.
• See how much autocomplete helps. Settings → Writing now shows accepted words, how often you use suggestions, their share of your typing, and estimated time saved. Charts show trends over time, with breakdowns by app category and model. Statistics start with this update and stay on your Mac as counts, without saving typed text, app names, sites, or window titles.
• MCP connections work with Claude Code and Claude Desktop. Fixed a connection timeout that prevented these clients from getting a response from lokalbot-cli mcp.
• Clearer Writing and Advanced settings. The autocomplete switch sits beside its model and permission status. Context, writing profile, health checks, and resource usage are easier to scan.
• Less clutter when completing actions. The persistent Undo bar has been removed. To reopen a completed action, go to Actions and choose All statuses.

The website also has clearer setup and switching guidance, new guides for using meeting notes with coding agents and Obsidian, and better mobile navigation.

Full changelog: v0.10.7...v0.10.8

LokalBot 0.10.7

What’s new

• Autocomplete can follow the conversation on screen. It can now read text from the window you’re typing in, including browser pages and chat apps that previously supplied little context. Nearby messages get priority, so suggestions can draw on what you’re replying to.
• More room for longer drafts. Autocomplete can keep more of your draft and the surrounding conversation together. It reads window text in the background and reuses unchanged context as you type.
• A new window-context setting. In Settings → Writing, Use the focused window as context is on by default, including after an upgrade. It works only when Autocomplete is enabled and Screen Recording permission is already granted. You can turn it off at any time.

Window text is read and used on your Mac. The captured image and recognized text are never saved or uploaded. App and site exclusions still apply, and private browser windows are skipped.

Full changelog: v0.10.6...v0.10.7

LokalBot 0.10.6

What’s new

• Use your ChatGPT or Claude plan. Choose either plan in Settings → Models → Think for Ask, meeting notes, and daily summaries, without an API key. ChatGPT uses browser sign-in; Claude uses your installed, signed-in Claude Code. Each plan asks for its own approval before receiving context, and usage-limit errors stop the request. These plans do not support Agent Mode.
• More responsive search while indexing. Ask and Quick Recall reuse the same query embedding across meeting and screen search. Background indexing works in smaller batches and gives typed queries priority.
• Keep screen search up to date. When semantic screen search is enabled, new screen text is indexed automatically in batches. Interrupted indexing resumes on launch, and failed passes retry.
• More reliable local models. Development and test copies use separate model-server ports and runtime files, preventing them from blocking the installed app. LokalBot also recovers model servers left behind by a terminated app.

Full changelog: v0.10.5...v0.10.6

LokalBot 0.10.5

What’s new

• Choose your Ask shortcut. Record a system-wide shortcut in Settings → General → Shortcut. LokalBot checks for conflicts with dictation and enabled macOS shortcuts, and pauses the current shortcut while you record a new one. The default remains ⌃⇧ Space.
• Faster Settings navigation. Switching Settings categories avoids repeated system login-item queries and unnecessary work in the Ask window. Localized labels also load more efficiently.

Full changelog: v0.10.4...v0.10.5

LokalBot 0.10.4

What’s new

• Better autocomplete in English and Serbian Latin. The hybrid decoder is now the standard whenever autocomplete is enabled, with improved partial-word completion and no extra experimental toggle. Gemma E2B remains the default local model.
• Less waiting for suggestions. The optimized hybrid reduced mean native generation time from 67.0 to 51.6 ms and p95 from 142.4 to 98.9 ms in the same-model, same-prompt benchmark. All 1,496 paired non-cancelled outputs matched. These are M4 Max measurements against the earlier hybrid; they exclude model loading, debounce, accessibility reads and overlay presentation. Benchmark and limits.
• More reliable suggestion placement. Improved on-screen caret detection, text-size matching and positioning keep ghost text aligned with what you type, including fixes tested in Chrome and Viber.
• Optional Fun-ASR-Nano transcription. Select a local Fun-ASR-Nano-2512 sherpa-onnx model folder in Models → Transcribe. LokalBot validates the folder and reads its weights in place; existing transcription choices are preserved.

Autocomplete preview

A 39-second preview with English and Serbian Latin before/after examples and the measured speed improvement. The typing scenes replay synthetic benchmark cases.

lokalbot-autocomplete-release.mp4

Explore the public interactive autocomplete tour.

After updating

Autocomplete remains optional and runs locally. Its improved decoder is used automatically when the feature is enabled. E2B stays the default, and your transcription model selection is preserved. Fun-ASR-Nano is an optional local model-folder choice.

Full changelog: v0.10.3...v0.10.4

LokalBot 0.10.3

What’s new

• Steadier dictation feedback. The recording overlay keeps its position and transcript space as dictation starts and finishes, follows light and dark appearance, and avoids flashing the preparation panel for short model checks. The microphone can start while LokalBot reads the original destination field.
• More consistent autocomplete. Slow caret lookup no longer holds up a ready suggestion, typing through a suggestion preserves its continuation, and window titles refresh when a composer is reused.
• Corrected Parakeet v3 model loading. Loading and download detection now use the encoder included in the pinned model files.
• Optional Parakeet Ultra transcription. Choose Parakeet Ultra 0.6B for local meetings and dictation, with a separate download of about 0.63 GB.

After updating

Existing transcription choices are preserved. Qwen3-ASR remains the default recommendation; Parakeet Ultra is an optional download.

Known limitations

Remaining autocomplete cases involving saved-fact freshness, conversation changes, and visual-caret capture scope are documented in the autocomplete review. This release includes the timing and continuation fixes and the window-title refresh, but does not resolve every finding in that review.

Full changelog: v0.10.2...v0.10.3

LokalBot for ChatGPT — preview

LokalBot for ChatGPT — preview companion

Requires Apple Silicon, macOS 15+, LokalBot, and Node.js 22.12+ (Node 24 LTS recommended). This is a foreground Terminal helper. The ChatGPT public listing is still pending; test connections require a client that supports custom remote MCP connections with OAuth.

This download includes a separate, compatible lokalbot-cli. It reads the same library under LokalBot's existing meeting-library permission. It does not replace or modify the installed app, enable permissions, or expose screen memory. Its source revision is recorded in SOURCE.json.

1. Open the signed DMG and copy the LokalBot-ChatGPT folder to a location you control, then eject the DMG. Keep the whole folder together.

2. Open Terminal in that copied folder. Pair your Mac:

   ./lokalbot-connect pair --relay https://mcp.lokalbot.com

3. Keep the connection running:

   ./lokalbot-connect run

4. Connect https://mcp.lokalbot.com/mcp in your client with OAuth. The consent page must be on mcp.lokalbot.com. Enter the pairing code there, never in chat. Check the requesting client and return domain before approving.

5. Enable Allow external agents to read your meeting library in LokalBot's privacy settings when ready to share requested results. Ask about a meeting or a saved commitment.

The Mac must stay awake and online, with the Terminal helper running. A code expires in ten minutes and works once; get another with ./lokalbot-connect code. Stop the connection with Control-C. Restart with ./lokalbot-connect run.

To revoke this Mac's connection:

./lokalbot-connect revoke

Disable meeting-library access in LokalBot and disconnect the client as well. Revocation cannot recall results already sent to ChatGPT.

The full library stays on the Mac, but requested meeting information passes through Cloudflare to your client. Read connection privacy, terms, and support.

Credentials are stored in ~/.config/lokalbot-chatgpt/device.json with owner-only permissions. Treat it like a password. For a separate synthetic test device, pass the same --config /absolute/path/device.json to every command and set LOKALBOT_STORAGE_ROOT to the test library before running. Do not share real meeting data, device files, or pairing codes in support reports.

This preview uses an unmerged source branch. It has no automatic updater; replace the complete companion folder when installing a newer verified preview. Do not bypass macOS security checks if an artifact fails verification.

LokalBot 0.10.2

What’s new

• More reliable dictation with Bluetooth headsets. Dictation handles microphone format changes and reconnects, shows when audio stops arriving, and keeps speech captured before an input failure.
• Dictation reaches the intended field more reliably. Replacing selected text, slower apps, cancellation, and consecutive dictations have improved delivery handling, with clearer clipboard fallback messages.
• Clearer recording and delivery feedback. The recording indicator waits for microphone audio, the level bars follow your voice, and an optional sound tells you when the microphone is ready. When a readable field does not show the pasted text, the overlay offers Copy.
• Choose how you dictate. Record a key combination or a modifier-only chord such as Control–Option, choose your microphone, or try Tap or hold in Settings → Dictation. Esc cancels a starting or active recording, and a warning explains when macOS Secure Input blocks the shortcut.
• Compose better preserves direct dictation. Ordinary questions and relayed sentences use a cleanup prompt that treats them as text to insert. Writing requests can use the newest focused-window text, and Last result shows which context contributed.
• Less repeated work for short Qwen dictations. Short recordings avoid an unnecessary second decode; recognition can retry without vocabulary hints when those hints overwhelm the result.
• Menu bar text follows macOS sizing at the default setting. Larger text choices remain available.

After updating

The microphone-ready sound is enabled by default and can be disabled in Settings → Dictation. Existing installs keep Option–Space as the default shortcut, and the microphone continues to follow the system input unless you choose another one. Optional Compose context grants remain unchanged.

Paste confirmation briefly checks only the original non-secure field and discards the text it reads. Unreadable fields and excluded apps are left unchecked.

Full changelog: v0.10.1...v0.10.2

LokalBot 0.10.1

What’s new

• Claude meeting notes and day digests work through Anthropic’s native API. Connections to api.anthropic.com now use the Messages API, fixing structured-output errors that prevented notes and digests from completing. Repeated requests over the same meeting or conversation can reuse Anthropic’s prompt cache, and Ask responses stream as they arrive.
• Choose English or Simplified Chinese for the interface. Settings → General → Appearance and first-run setup offer Follow System, English and 简体中文. The selection updates windows, app-owned menus and settings search immediately, while transcription language, generated notes and saved content keep their own settings.
• A smaller on-device model for search by meaning. EmbeddingGemma 2 replaces Harrier, with a smaller download and lower idle memory use. It improved Serbian and Montenegrin meeting queries in our evaluation, including queries over English meetings. The bundled llama.cpp runtime is updated to support it.
• Find a speaker’s name as you type. Suggested names can be searched by name or email, regardless of word order, case or accents, including common spellings of Serbian names.
• Day digest controls are easier to use. Ask About This Day, Write Digest and the digest actions menu use the regular control and text sizes.

After updating

Search by meaning downloads the new model (about 310 MB) and rebuilds existing search vectors on your Mac. Keyword search remains available during the rebuild. After the new model has been verified, obsolete embedding-model files are removed to reclaim space.

Full changelog: v0.10.0...v0.10.1

LokalBot 0.10.0

What’s new

• Meeting notes adapt to the model you use. Notes are now split to fit the selected model’s real context window instead of one small size for every model. The window comes from OpenRouter’s live model data, from servers that report it (vLLM, Groq, Together, Mistral, LM Studio and llama-server on your Mac), or from Cerebras’s and OpenAI’s documentation, and each request stays within the provider’s structured-output limits. A 20-minute meeting now takes one or two requests on hosted models such as Qwen 3.8 27B, GPT-5.6 Luna and GLM 5.3 Flash, instead of four.
• Stuck summaries recover. A summary part that ran out of output before writing anything is read again from the start instead of being asked to continue from nothing, and Summarize again retries a part that an earlier run had stopped. On Automatic reasoning, notes no longer spend their whole answer on hidden reasoning with models that think by default.
• Choose a reasoning level per model. Settings → Models and the Agent composer offer the reasoning levels the selected model accepts, from Off to Max. Automatic keeps each task’s own setting.
• Day digests stay put. Re-summarizing a meeting, editing its transcript, ticking an action item, saving a moment or new coding-agent activity no longer deletes that day’s digest. It stays visible and is marked out of date until it refreshes. Digests also outlast the screen-text retention window, like overnight reviews. Deleting a meeting, a capture or saved agent sessions still removes the digest built from them.
• Fewer wrong autocomplete suggestions. A suggestion is held back when the model is unsure of its first word. New installs use the visible text above the field for better suggestions, and private windows are never read. Autocomplete stays quiet in open and save dialogs, clears a suggestion left alone for a minute, and points out when macOS’s own inline predictions are on.
• Open a meeting in Claude or Codex. The meeting’s More menu starts a Claude or Codex conversation with the meeting’s title, summary and transcript already typed in; nothing is sent until you press Send. It can also copy the meeting folder’s path, and workspace menus use the standard menu font.
• Agent Mode asks before every tool except reading files in your workspace, including tools from MCP servers, unless you choose Full access.
• The main window resizes properly on macOS 26. Sections with a split view can shrink to their real minimum width again.
• Updated on-device runtimes. MLX Swift 0.32.3 and MLX-LM 3.32.3 for on-device speech, plus updated networking and model-loading dependencies.

Full changelog: v0.9.6...v0.10.0