Releases

My node on iPhone too, what your machine did at a glance, recipes that run on your node, and the team of agents your machine can now run.

My node on iOS

• Pair your computer in Settings › Providers › + › My node (Invincible). Paste the code, or open the invincible://pair link from the camera.
• Chats, signed receipts and "phone first, then my node" all work as on Android.
• The agent screen. Remote › Agents · my node: tasks on that machine, approval cards, and scheduled tasks.

What your machine did (Android and iOS)

The list button on the agent screen opens the node's Activity, read from its own logs:

• Requests: every request the node answered: who asked, which model, tokens, and whether anything left. Each receipt is checked against the key pinned at pairing.
• Actions: what the agent ran, and who allowed it.
• Memory and skills: what the agent remembered (you can make it forget), and your skills.

Recipes on your node

• Runs on in the recipe editor: this phone, or your node.
• On the node, a recipe runs on its schedule, always on and with no 8-minute limit, on the node's models, with only the recipe's tool groups (web included).
• Your time, wherever the node is. "07:00" means 07:00 where the phone is.
• Results come back as chats in the Recipes folder, with the node's signed receipts and a notification.

A team on your machine

• Team members. When the node's agent has a team, each member's answer shows under its name.
• You know who's asking. An approval says which member wants it ("coder: Run a command?").
• New approval cards for web fetches and MCP tools.

Fixes

• Several CI tests that depended on timing are steadier.

Your own computer as Autobot's next tier: pair a machine running Invincible, chat with its models, hand its agent tasks you approve from the phone, and check what it signed.

My node (Android)

• Pair a computer. Install Invincible on a desktop or VPS (curl -fsSL https://raw.githubusercontent.com/wckdboy/invincible/main/install.sh | sh), run invincible pair, and scan the code with the camera app (or paste it). Settings › Providers › My node (Invincible). No account, no relay: the phone and the node pin each other's keys (Noise).
• Chat with its models like any provider. Connections follow the kill switch (offline refuses them; Tor and SOCKS5 carry them with remote DNS) and show in the Privacy Center's audit log.
• Signed receipts. Every answer from the node is checked against the receipt key pinned when you paired: the receipt line says "kept on studio ✓" when nothing left the node, "studio → 1 host" when it used a cloud model (with the hosts, routes and what it redacted), or "studio unverified".
• Phone first, then my node. With "phone first" on a node chat, this phone answers first and a weak answer goes to your node instead of a cloud model. The node is hardware you own, so a sensitive prompt may go there and "Never" allows it; these turns are marked private, so the node never sends them on to a cloud model.

The agent on your machine (Android)

• REMOTE › Agents · my node. Give the Autobot Agent running on that machine a task: shell commands and file work, on its models. Every command, file write and read outside its working folder comes to the phone as a card first: Allow, Always (this conversation) or Deny.
• Scheduled tasks. The calendar button lists what the agent runs on a schedule (next run, how the last one went), adds tasks (cron, @daily, every 30m) and removes them. Scheduled runs only do what you allowlisted.

Fixes

• Node chats no longer show a REDACT chip: nothing is redacted for hardware you own.
• The streaming check with a real node leaves room for thinking models.

iOS: the pairing and receipt code is shared and tested on CryptoKit; the iOS node connector comes next, so "My node" says "not on iOS yet".

Images come out the shape you pick, a full-screen viewer, recipes that run on their own, phone-first answers and a panic PIN.

Images (both platforms)

• 16:9 means 16:9. Picking a wide or tall shape could still give a square: on-device models compiled for one size (Core ML on iOS, QNN on Android) quietly swapped your size for theirs. They now render their square and return your shape cut from it, and Tune says "compiled for 1024² · other shapes are cut from the square".
• The resolution is a pixel budget. 16:9 at 1024 is about a million pixels in that shape (1368×768), so wide images keep the detail of the square. Every shape chip names the exact size it makes, and the picker remembers your choice instead of snapping back to 512.
• Picking a model keeps your shape instead of resetting to its square. MNN sizes round to the nearest 64 instead of always down.
• A server that sends back a square anyway is cut to your shape (Android) or named in a notice (iOS), and the gallery records the size it really got.
• Full-screen viewer. Tap a result, a gallery image or an image in chat: pinch or double-tap to zoom, pan, swipe through the batch, swipe down to close. The info sheet shows the real pixel size and every setting used.
• Do more with an image: use it for img2img (prompt and settings carried over, strength showing), inpaint, vary, reuse settings, same seed, upscale 4× (when an upscaler is installed), share, save, copy settings, favorite, delete.
• The agent's generate_image takes an aspect ratio and keeps the shape when given one side.
• img2img without an input image now says so instead of quietly running as text-to-image; Tune says when a model can't apply your LoRAs.

Recipes (both platforms)

• On-device agents on a schedule. Settings → Recipes: a prompt, a schedule, a model and the tools it may use. Runs in the background on the phone (WorkManager on Android, background processing on iOS), once for a missed time, safe across daylight-saving changes.
• A private notification opens the result chat after the app lock. While the phone is locked, iOS says "Unlock to run <name>". Recipe data is encrypted and goes with a wipe.

Phone first (both platforms)

• The phone answers first. Off by default. Settings → Privacy → Phone first (ask / auto / never) and a chip per chat: in a cloud chat the installed on-device model answers, and

Cleaner prints, a faster start, and dictation that keeps up: a polish release.

Cleaner ink (both platforms)

• No more doubled dots. On Android, the paper behind the Image screen, chat, voice, Models, Gallery and the lock drew every dot twice, with a faint copy a few pixels off. Paper is one plate again.
• Prints land on their own dots. When a result printed, its colour plate started half a dot away and read as a second grid. It now starts as a thin rim on the same dots and settles.
• Pictures print in where they show. A wide image in a square chat thumbnail or a Studio chip no longer prints a stretched copy next to the cropped one.
• One dot field per screen: the voice panel, the empty Models search and empty chat no longer stack a second field under the first, and image proofs sit on whole pixels.
• iOS: a halftone preview always matches its own image. A new image could pick up an old image's dots.

Faster (both platforms)

• Android starts sooner. The app now ships a Baseline Profile, so startup code runs precompiled even on a sideloaded install. On the emulator: 13–17% less time to the first real screen and about 20% less main-thread work.
• iOS starts sooner. Fonts and device checks load in the background, about 90 ms earlier to the first frame in the simulator.
• Live dictation shows words sooner and more often. Partial passes only encode the audio heard so far, 2–4× less work. Stopping usually has nothing left to transcribe.
• Faster draft-model checks. Each speculative step costs less, a draft that failed to load is retried later, and a failed check carries on without the draft instead of ending the answer.
• Image model warm-up. It loads again ahead of Generate after it was freed, switching models no longer loads both, and on Android it stays loaded while the Image screen is open.
• Smoother lists: the iOS chat list scrolls with fewer hitches, and Android's Models list only redraws rows that changed. Catalog previews and parsed chat text have a memory limit on iOS.

Fixed

• Read-aloud no longer cuts off the last words before the assistant uses a tool.
• Talking over a spoken answer no longer triggers by itself when the voice resumes after a pause.
• iOS: the last words you say before tapping stop are no longer lost from dictation.
• A quick double tap on the mic starts one dictation, not two.
• The Image scree

Your words stay yours in the cloud, replies come faster on the device, voice answers while it thinks, and you can see what every model costs.

Redact before cloud (both platforms)

• Placeholders, not your data. Before a chat goes to a cloud model, names, emails, phone numbers, addresses, card, IBAN and account numbers, and your memory facts become placeholders like [PERSON_1]. The answer comes back with your real words, put back on your phone.
• Tools still work. The agent gets the real values back on the phone, so it can still write the email or fill in the file.
• On by default, in Settings → Privacy & security. A REDACT chip turns it off for one chat. "This phone" and local servers are never redacted.
• On the receipt. It shows what was hidden ("2 names · 1 email"), never the values. Code, links, dates and version numbers are left alone, so coding chats keep working.

Faster replies with a draft model (both platforms)

• A small model guesses, your model checks. Install Qwen3 0.6B next to Qwen3 4B Instruct 2507, Llama 3.2 1B next to 3B, or Qwen2.5 Coder 0.5B next to the Coders. On a Mac GPU, code, lists and rewrites came 1.3–1.7× faster and long answers about 2×.
• Every word is still your model's. Thinking budgets, tool calls and the stop button work as before. Short agent steps skip the draft, so they never get slower.
• Auto / On / Off in Compute. Auto drafts only when the pair fits your memory comfortably, and pauses when drafting doesn't pay. The benchmark and the diagnostics report show the speedup and how many guesses were kept.
• Qwen3.5 models don't use drafts: their memory can't take back wrong guesses cheaply. A draft from another model family is refused with a reason.

Voice that keeps up (both platforms)

• Your words as you speak. Dictation shows settled words in ink, with the rest firming up as you go. Audio is never saved.
• Answers start out loud after the first sentence, not after the whole reply. Code is skipped; numbers and abbreviations are read whole.
• Just talk to interrupt. In hands-free mode, speaking over an answer stops it, keeps your first words and listens.
• Stopping dictation is faster: the last pass starts right away.

Usage, counted on your phone (both platforms)

• Settings › Usage: tokens, requests and cost for today, this week and this month, a bar per day, and a

Faster and smarter on-device chat: a real thinking switch with a budget, the NPU reads your prompt while the CPU writes the answer, proper formatting in answers, and a lighter Run screen.

Thinking you control (both platforms)

• A real switch. The ∴ think chip above the composer cycles Auto / On / Off; long-press for a budget (256 to 4k tokens). It's remembered per chat, with optional per-model defaults. Thinking is turned off through the model's own chat template, not a prompt the model can ignore.
• A budget the engine enforces. When thinking reaches its budget, the engine wraps up and the model answers. Part of the output is always reserved for the answer, and a loop guard ends thinking that goes in circles.
• Auto is sensible. Small models (under 4B) skip thinking in quick chat, which is faster and, on 0.8B models, also more accurate. Agent and code modes think within a budget. Subagents and summaries never think.
• Before and after: Qwen3.5 0.8B used to think about 2,000 tokens and never answer. Now it answers in 25 ms with thinking off, 1 s with a 128-token budget, and 3.3 s on Auto in agent mode.
• Cloud providers get their real reasoning control (OpenAI, Claude, OpenRouter, Qwen, DeepSeek, Z.ai, Groq, Gemini, vLLM, llama.cpp server). Auto leaves each provider's default alone.

NPU reads, CPU writes (Android, Snapdragon)

• A new npu+cpu setup: the Hexagon NPU reads your prompt, then the CPU writes the reply. On a OnePlus 15 with Qwen3.5 0.8B: prompts read at about 2,500 tok/s (CPU about 100), and replies are written at about 50 tok/s (NPU alone about 40). A long agent prompt that took 10–20 s to read now takes under a second.
• Auto now ranks engines by a typical turn (about 1,000 prompt tokens + 200 generated), so it picks npu+cpu where that's fastest. If memory is short, it falls back to the NPU alone.
• The Compute screen shows all five engines. An on-device bench (engine/llama/bench/android.sh) is included for anyone measuring their own phone.

Formatted answers (both platforms)

• Answers render headings, bold, italic, lists (nested and checkboxes), quotes, tables, code blocks with copy, and links.
• Still private: nothing in a message can make a network request. Images show as text ("[image: alt · host]"), HTML stays literal, and links ask before opening in your browser, with a warning for plain http

Clear Settings and a new About screen, honest engine labels, the right image defaults, and smoother screens. Tuned on a real OnePlus 15.

Settings and About (both platforms)

• Settings is laid out in one clear order: Privacy & security › Agent › Appearance › Data › About. Every setting has exactly one home. Providers, image backends and accounts live in Remote, and Panic wipe lives only in Settings.
• About shows your version, build number and build date, and this release's "What's new". The notes are built into the app, so they work offline. "All releases" and "Source code" open in your browser. It also has licenses, the privacy statement ("No analytics. No telemetry. Network only where you allow it.") and Copy diagnostics.
• Android: About shows the release signing certificate and checks it against the APK you're running, so you can verify your install is genuine.

Labels you can trust

• Route tags show the engine actually doing the work.
• Android: NPU, GPU or CPU, for images and for chat (from what the engine reports).
• iPhone: Neural Engine or GPU for images, GPU (Metal) for chat.
• "CPU" no longer appears for models running elsewhere.

Better image defaults

• Picking a model applies its recommended steps, guidance, sampler and size, without overriding anything you changed. Illustrious DMD2 now opens at 8 steps and 1024² instead of 20 steps at 512². SDXL models never default to 512².

Cleaner screens

• Run: the status line fits, the card grid is aligned, the image art is clean, subtitles wrap instead of cutting off, and an empty "Recent runs" is hidden.
• Models: the phone header fits; installed rows read like "image · NPU · 4.5 GB".
• No halftone dots behind text or numbers anywhere.
• iPhone: nothing is cut off on the iPhone 17e or at extra-large text, and the 17e shows its chip name.

Smoother

• Android: halftone fields are drawn as batched dots. On the OnePlus 15, Imagine's dropped frames halved (22% → 11%), and slow frames went from 42 ms to 16 ms (p90).
• iPhone: screens hidden under the Studio stop animating (stutters 3–6 → 0). Studio Library reloads thumbnails 4× less often while scrolling, and Imagine at rest uses half the main-thread time.

Measured on a OnePlus 15 (Snapdragon 8 Elite Gen 5)

• Qwen3.5 0.8B: CPU 49 tok/s generation, NPU 576 tok/s prompt reading. After the Compute benchmark,

A calmer Android app: much less text, fewer taps, and the same Settings layout as iPhone.

Less text, fewer taps (Android)

• Settings is grouped into Connections / Agent / Appearance / Security / Data / About. Rows show their state on the right (providers "1", Privacy "Direct", Vault "Off") instead of a subtitle. Explanations are behind ⓘ.
• Chat with a downloaded model: 2 taps instead of 3. A new chat opens with the keyboard up. Tapping Chat again reuses an empty chat instead of piling up "New chat" rows.
• Generate an image: 2 taps instead of 3. The prompt is focused, and the Generate bar now sits above the keyboard instead of under it.
• Shorter everywhere. Chat rows show the model's name ("Qwen3.5 0.8B") and "just now". Model rows drop licence and footnote text, and download progress shows once, in the panel. Remote, Privacy Center, Vault, Imagine and Studio lose their paragraphs: one short line and one button where something needs doing.

Look

• Living cards match iPhone exactly: a fixed 320 dp wave at 0.45 rad, period 12 s. Every dot stays visible at the wave's low point.
• Run screen cards fade their dots under the title and model name, so the text stays legible.

Files

• autobot-1.6.4-release.apk: signed release (ai.wckd.autobot, arm64-v8a, Android 14+). Signer certificate SHA-256 9a1378acf9587885e7920355558c85a1ec0e89304f58c153feebf90fdee7f3f1.
• autobot-1.6.4-debug.apk
• autobot-1.6.4-mapping.txt: R8 mapping, for reading crash reports
• SHA256SUMS.txt

iOS: TestFlight build 1.6.4 (no iOS changes since 1.6.3).

On-device chat is fixed and much faster: no more failures every other turn, follow-up replies start almost at once, and Stop takes effect immediately.

Fixed (both platforms)

• The default on-device chat models failed every other turn. Qwen3.5 0.8B, 2B and 4B use recurrent layers whose memory can't be rewound to an arbitrary point. The engine ignored that, so every second message ended in "DECODE_FAILED", and the next one re-read the whole conversation. The engine now keeps snapshots of that memory (after the system prompt, and near the end of each prompt) and resumes from the newest usable one. Replies are identical to a fresh run.

Faster

• Follow-up replies start in well under a second. The conversation already processed is reused, so the model reads only your new message. On a Mac GPU with a 3,000-token agent prompt, the second turn now starts in about 0.1 s instead of failing or re-reading for 1–2 s. On CPU, 5–8 s per turn becomes 0.1–0.5 s.
• New chats with the same agent reuse its system prompt and tools, so the first reply comes sooner too.
• The model loads when you open a chat, while you type, instead of when you press Send. It holds off while an image, audio or dictation job runs.
• Text appears sooner: tokens are shown as soon as they're sampled, and the first words up to two tokens earlier.

Stop means stop

• Stop while the model reads your prompt now takes effect in milliseconds (it could take over 7 s on CPU).
• Stop during model loading cancels the load.
• Android: a Stop pressed while a reply was waiting or loading could be lost, so the phone generated a full answer for nobody. A finished reply's Stop could also cancel the next one. Each reply is now stopped on its own.

Files

• autobot-1.6.3-release.apk: signed release (ai.wckd.autobot, arm64-v8a, Android 14+). Signer certificate SHA-256 9a1378acf9587885e7920355558c85a1ec0e89304f58c153feebf90fdee7f3f1.
• autobot-1.6.3-debug.apk
• autobot-1.6.3-mapping.txt: R8 mapping, for reading crash reports
• SHA256SUMS.txt

iOS: TestFlight build 1.6.3.

On-device image generation fixed for Local Dream NPU models, sturdier downloads and engines, safer agents, pixel-exact dots, and much less text on iPhone.

Fixed: image generation on Snapdragon (Android)

• NPU image models now run. A phone whose only image models were Local Dream / xororz NPU packages (SD 1.5, SDXL, Illustrious and others) never got a "This phone" image backend, so none of them could be used. They now appear in Imagine and Studio.
• Builds picked like Local Dream does. Snapdragon 8 Elite Gen 5 now gets the 8gen2 builds (SDXL: 8gen3), matching Local Dream. If you installed an NPU upscaler before, reinstall it so you get the right build.
• SDXL uses less memory. Only the model in use stays loaded. SDXL frees its text encoders before loading the UNet, and the VAE encoder only loads for img2img and inpainting.
• Clear errors. A failure now names the stage (loading, text encoder, UNet) and quotes the engine's own message. Models › Compute › Report includes a live NPU check: how far the NPU setup got, the runtime version, the chip's library, and a test load of your smallest NPU model. If something still fails, copy that report and send it.

Downloads and engines (Android)

• A cancelled download no longer comes back after the app is killed. Pause and cancel take effect at once, even on a stalled connection.
• Resuming a file that changed on the server restarts it cleanly instead of splicing two versions together.
• A disk that fills up pauses with a storage message. Gated Hugging Face models tell you to sign in and accept the terms.
• Models whose files were deleted outside the app are detected and offered for re-download (only the missing files).
• A crashed engine is released and restarted cleanly. Cancelling during model load now works. Broken or wrong model files are named before loading.
• Tapping a Whisper model opens voice mode.

Agents (both platforms)

• Every tool that can reach the network is checked for private data, and an approval prompt can no longer override a refusal.
• Approvals cancelled midway are logged correctly. Replies the model cut short are no longer retried as empty.
• Privacy receipts list every model your data went to, including subagents and summaries.
• Studio remote runs time out and cancel properly.

Look and feel

• Pixel-exact dots (both platforms). Bars, loaders, living cards and bac

A reliability release for iPhone: no more waiting forever, faster Stable Diffusion, and clear messages whenever something is set up wrong.

Fixed (iPhone)

• Panic wipe could crash partway through. Shutting down the agent called back on the wrong thread and stopped the app. Fixed.
• No more endless waits. Two places that could wait forever after a cancelled download or request now always finish. Testing a provider or an MCP server now gives up after 30 s, and testing a ComfyUI server after 20 s, with "No answer from <host>" instead of a long spinner.
• Stable Diffusion 2.1 is faster. On-device image generation now uses the backend your Compute benchmark picked instead of ignoring it. Memory settings are tuned per device: models stay loaded on 8 GB+ phones and load stage by stage on smaller ones.
• Multi-file downloads no longer re-download finished files after a later file fails, and a disk that fills up mid-download is reported as a storage problem.
• Tells you what's wrong.
• The chat shows a banner with a fix button when there's no provider, Offline mode is on, a cloud provider has no API key, or the on-device model is missing.
• Hugging Face and Civitai search say why a list is empty: Offline mode (with a shortcut to Privacy Center), a refused token, an error, or no results.
• Missing voice and audio models say where to get them.
• Counts read properly ("1 session", not "1 SESSIONS"), model rows show one size, and Studio number fields keep what you type ("-", "1.", or clearing the field).
• Stopping read-aloud also stops the sentence being synthesised.

Under the hood

• 58 new iOS tests (371 app tests, 51 engine tests) covering downloads, running models, the agent bridge, and error and empty states. Every async test now has a time limit.

Files

• autobot-1.6.1-release.apk: signed release (ai.wckd.autobot, arm64-v8a, Android 14+). Signer certificate SHA-256 9a1378acf9587885e7920355558c85a1ec0e89304f58c153feebf90fdee7f3f1. No Android changes since 1.6.0. Android fixes come in the next release.
• autobot-1.6.1-debug.apk
• autobot-1.6.1-mapping.txt: R8 mapping, for reading crash reports
• SHA256SUMS.txt

iOS: TestFlight build 1.6.1.

A Creative Studio for images, audio and video, a calmer and lighter motion language, and a full performance and bug pass so the app feels quick again on both platforms.

New

• Creative Studio. A focused mode for making things. It has rooms for image, audio and video, a job queue that keeps going while you work, and a library of everything you've made with its prompts and settings. You get more control: seed, steps, guidance and size, plus reusable prompt snippets.
• On-device audio. Stable Audio Open Small generates sound effects and short music clips on the phone. Powered by Stability AI.
• Workflows. Connect a ComfyUI server you run yourself (the connection goes through the kill switch and the network audit). Autobot detects each workflow's inputs and turns them into simple forms. The agent can also run them through a new run_workflow tool.
• Ink motion v3.
• Progress bars are now a row of dots. An empty bar is empty, and dots grow bigger as it fills.
• "Working" is a travelling wave of swelling dots.
• Hero cards and empty states sit on a slow, gently breathing halftone field.
• The animations use one shared clock per tier, are drawn only (they never trigger a re-render), stop when idle or on Low Power/Battery Saver, and turn off with Reduce Motion.

Faster

• iPhone:
• A streaming frame now takes 1.2 ms instead of 5.6 ms.
• The data sent for each streamed update dropped from about 520 KB to under 1 KB, because only the new text is sent.
• Chat state is split, the message list is memoized, and storage work runs in the background.
• Android:
• A streaming chat recomposes once instead of on every token.
• Frame time dropped from 42 ms to 24.5 ms.
• Session projection is incremental, and providers are cached.
• MCP connections pause while the app is in the background.
• Both: at most four chats keep a live agent in memory. Approvals are matched exactly to the tool call that asked for them.

Privacy

• ONNX Runtime telemetry removed (Android). The ONNX Runtime library added a component that sent usage pings around the kill switch and loaded a 34 MB library into every process at startup. Autobot now strips it out.

Smaller

• The Android release APK is code-minified with R8: 80.6 MB instead of 110.5 MB. A mapping file ships with each release so crash reports can still be read.

Fixed

• Android

Privacy you can see, and a much sturdier app. Every answer now carries a receipt, private data can't slip out through tools, your documents become searchable on the phone, and a full crash and freeze review is fixed on both platforms.

New

• Privacy receipts. Under every answer: "on device", or which hosts were contacted, bytes out, route and the model's data policy. Tap for the full receipt. Stored with the chat, never sent anywhere.
• Taint tracking. If the agent tries to send something private (memory notes, file contents, numbers from your images) through a web fetch, search or MCP tool, it must ask — even in full-access mode — and the card highlights exactly what would leave.
• Private vault (Settings › Vault). Add notes, text, PDFs (and photos with on-device OCR on iPhone); they're embedded on the phone (Qwen3-Embedding, llama.cpp), encrypted and searchable by you and by the agent. Never uploaded. Off by default.
• Undo for agents. Every file change the agent makes shows a diff and can be undone per step or per turn (and redone). A "Files changed" sheet lists everything.
• Chat organisation. Rename, pin, folders, archive, search inside messages, swipe to delete with undo, multi-select, sorting and date groups. Folders and pins travel in .autobot backups.
• Ask Autobot from anywhere. iPhone: Shortcuts/Siri, Action button, share sheet, widgets, Control Center voice control. Android: share target, text-selection "Ask Autobot", Quick Settings voice tile, launcher shortcuts, widget. All behind the app lock; only your own Shortcut can send a prompt.
• Downloads that keep going. Android uses user-initiated transfer jobs (survive leaving or swiping the app, no 6-hour limit); iPhone relaunches in the background to verify and continue the queue. New downloads panel with speed, ETA, queue and why something waits; optional Wi-Fi only (Android).
• Ink motion v2. Ink-dot loaders everywhere, a progress bar with determinate / sweeping / paused / failed states, a quiet drifting halftone paper behind empty screens and the lock screen, out-of-register screen transitions and dissolving rows. Light on battery (≤30 fps, 10 fps ambient, stops when idle) and off with Reduce Motion.

Fixed

• A chat could crash every time it was opened after an on-device model reused a tool-call id (both apps).
• Image generation, chat, speech and voice m

Talk to it, hear it, watch it print. Voice mode, a neural voice that runs on the phone, a new motion language, and OpenCode Go support.

New

• Voice mode (VOICE in the chat bar). Tap to talk: whisper transcribes on the phone and sends; the answer is read aloud when the turn ends. Tap while it speaks to interrupt. Optional hands-free mode stops listening when you go quiet. Approvals are announced, never auto-approved.
• Kokoro neural voice (Models › voice). A natural-sounding voice that runs entirely on the phone: ONNX on Android, Core ML on iPhone. Pronunciation uses Kokoro's own open dictionaries; nothing is sent anywhere. Without it, read-aloud uses the phone's offline voices only.
• Ink motion. While an image generates, the halftone plate flows and fills; previews develop in along the flow; the finished image prints in with the red plate settling into register. Steps stamp in, the approval card arrives out of register, thinking shows as rippling dots. Capped at 30 fps, paused when idle or off screen, and off with Reduce Motion / animations disabled.
• OpenCode Go as a provider: one key, ~30 models. Each model shows its data policy (no retention, kept N days, or trains on your prompts — those need confirmation and are never picked automatically). A helper model per provider runs subagents and summaries to save quota. When a 5-hour or weekly limit is used up, Autobot says when it resets, and can fall back to the helper or on-device model if you allow it.

Fixed

• Android read-aloud could use a network voice; it now only uses voices installed on the phone.
• iPhone read-aloud was silent right after dictation.

Files

• autobot-1.4.2-release.apk: signed release (ai.wckd.autobot, arm64-v8a, Android 14+). Signer certificate SHA-256 9a1378acf9587885e7920355558c85a1ec0e89304f58c153feebf90fdee7f3f1.
• autobot-1.4.2-debug.apk
• SHA256SUMS.txt

iOS: TestFlight build 1.4.2 (10).