nanoMuse

nanoMuse 1.0.8 · Crew

🐉 nanoMuse 1.0.8 “Crew”: four days after Logbook, your devices keep each other posted. A scheduled task on the computer that needs a yes asks on your phone; every device reads the step another is on and says when it is waiting for you; the wait before a first token is spoken of, with your place in the queue; a thumb on a reply or the reason for a Stop goes to the account's record as one word, and the operator reads the record question by question. The phone no longer crashes when you stop talking, the iPhone no longer stops your music, the desktop's Library takes documents of your own, and the runtime behaves on Windows.

1.0.7 went out on 2026-10-25. Since then seventy-two pull requests were merged, and this version is the fifth look at the whole project, with one change at its centre: the devices keep each other posted. Until now a task that ran on one device was a closed box to the others: the phone saw Mac is working… until the reply came, a scheduled task on the computer that hit a step needing a yes waited for someone at the keyboard, and a person at the back of the relay's queue saw the typing dots and nothing else. Now a scheduled run on nanoMuse Desktop sends its permission card to the account's phones, payment steps included, with Allow once, Always in this task and Deny; the first answer on the computer or on any phone decides and the other cards close, and the grant stays on the computer (#652, #672 by @cutey-none, with the grants per chat and per task of #642). Every tool call tells the relay the step, in the model's own words when it gave them, so the line under another device's prompt reads Pixel 8 is working: 打开携程网站 and Pixel 8 is waiting for you while a card holds its run (#617 by @z1010223zero-pixel, #635, for #610). The relay tells a streamed turn where it stands while it waits for a provider slot and keeps the line alive while the provider thinks, and the apps draw the wait under the typing dots: The model is busy, 15 s so far with Cancel, or 3 ahead of you; the phones also compact a conversation at 32k tokens instead of 108k and carry about an eighth less in every call, since the time to the first token grew with the prompt (#626). Around the record that 1.0.7 opened: every call says who started it (a person, a scheduled task, the feed, a goal's check, another device) and where it sits in the conversation, a turn on a key of your own joins the training view when Help improve is on, what happened around a turn reaches the relay as codes and never as words (what the Sentinel decided, how an approval went, why a reply was stopped, a thumb up or down), and the operator's console counts the record by questions rather than calls and lays each question out once (#653, #654, #655, #660, #661, #662, #663); its figures read an index instead of every body, which on the hosted relay's 30 GB turned half a minute into a third of a second (#680, #681). The rest is what a check and the people using it found. @Yicheng-Mark made the runtime at home on Windows: a private file stays its owner's, a config.toml Notepad or PowerShell wrote is read, a file is written with the lines the

nanoMuse 1.0.7 · Logbook

🐉 nanoMuse 1.0.7 “Logbook”: two days after Beacon, the relay keeps the record of every conversation with the account's models, screenshots included, and the apps add the turns they ran on a key of your own, so your devices share one history whichever of them, and whichever key, answered; the drawn face's clips travel between devices; nanoMuse Cloud answers the nanoMuse apps only; the operator's console is rebuilt around what the model endpoint is doing; and Data controls on every app says what is held and deletes it with one tap.

1.0.6 went out on 2026-10-23. Since then twenty-seven pull requests were merged, and this version is the fourth look at the whole project, with one change at its centre: the record. From relay 0.29 nanoMuse Cloud keeps each chat turn it answered, in full: the request as it sent it to the model (the system prompt with your memory and SOUL, every message, what tools returned, the screenshots of a screen operation) and the answer as the model gave it; the pictures are stored once each as files named by their hash, so a screenshot an app resends on every step is one file. It is the account's own data: your devices read it back, the person running the relay can see a problem whole, nobody else can, and one tap under Data controls removes all of it. The 1.0.7 apps take part from both sides: each chat call names the device and the conversation it belongs to, and a turn answered on a key of your own goes up afterwards with its pictures, never the key or the address, so the record is whole wherever the model was (#579, #591). Around it, three more things the account does: the face's clips travel with the face, so a look drawn and animated on one device is worn and moves on the others without a second rendering job; each device says in its hello which provider and model answer its chats (the two ids, nothing else), so the Devices list and the operator's Platforms page can say what people run; and the models answer the nanoMuse apps only (relay 0.28): a pool worker, an SDK or a script on an account key is turned away with 403 client_not_supported and a sentence pointing to a key of its own, while every app of 1.0.6 and earlier goes on unchanged (#570, #573). The operator's console at /app/admin is a new application, in 简体中文 and English, with a Live page that reads the provider gate every three seconds and pages for platforms, retention, screen operation and data quality; the relay writes one row per chat call saying how it went, never what was said, and its heavy pages read on a second connection off the requests' path, so a month of the ledger costs the operator the wait and nobody else anything. The rest is the kind of thing a check finds: the runtime names itself on every model call, treats a remote MCP server as sending data out, asks before telling a coding agent what to do, keeps a webhook's key out of the transcript and shares connector addresses without their secrets; the desktop keeps a failed push's lines for the next one; the phone stops the account's work before signing out; the iPhone keeps a second account's profile and files apart from the first. Beside the check, @mongo027 made a thinking model's reasoning show while it thinks (#484); @cutey-none added Parallel as a web search provider (#566) and fixed the desktop settings' tint (#569); @Mola-maker let one scheduled task fire at several times a day (#558); @hobostay made the desktop check who stops or approves a remote task (#567); @z1010223zero-pixel

nanoMuse 1.0.6 · Beacon

🐉 nanoMuse 1.0.6 “Beacon”: nine days after Lookout, a third look at the whole project with the forty-eight pull requests that arrived in between; nanoMuse Cloud serves one multimodal model and leaves pictures and clips to a provider of your own, the relay's sign-in gives nothing away, the runtime finishes what a restart interrupted, the desktop asks before a restart, and the phone reads the Feed posts you liked.

1.0.5 went out on 2026-10-14. Since then forty-eight pull requests were merged, more than in any week before, and this version reads the project through once more with them in: the relay, the Python runtime, the web console, the desktop, the Android and iPhone apps, the docs and the workflows, area by area. The largest change is to the account: nanoMuse Cloud now serves one model, Qwen/Qwen3.6-27B-FP8, which reads text and pictures and does both the chat and the operating of a screen, and it no longer pays for image or video generation (#536, #539, #545, #546). A drawn face, a picture asked for in a chat and the face's clips come from a provider of your own on every client, the pickers list only sources that can, and an empty row names who could; the default dragon and the four new preset looks are pictures already made, so a look is yours without a key. The relay is 0.27.0: a wrong password and an unknown account get the same answer, every limit is decided in one transaction, a Hub credential speaks for one device, and an invite bonus waits for real use (#465). The rest is the kind of thing a check finds: a reminder or a routine the runtime did not get to finish runs at the next start; the browser tool refuses local and private addresses as web_fetch always has; a streamed reply is retried only before its first word; restarting the desktop's Host asks first, as quitting does; a face drawn with your own key while signed out is worn rather than refused (#521); a remembered answer to an approval relayed from another device covers that step, not every step that starts the same way; hidden steps on the iPhone stay hidden after a reset (#548). Beside the check, @Mola-maker made forgetting a memory visible and undoable and gave the web console's settings a search (#533, #541) and fixed nine things on Android and the desktop; @jiangkaiqi2005 taught the Windows desktop to download its updates in the background and install them on quit (#458, #459, #460); @cutey-none's @ menu lists your other devices on the iPhone and a task sent to the desktop streams its text back (#506, #513, #515); @Battleplus added a fourth trigger kind and the data download (#487, #489); @Yicheng-1218 let each avatar mood carry a clip of your own (#427, #481); @isHeSatoshi put a search in the desktop's settings (#523); @hobostay kept a trimmed catalogue trimmed (#527, #530); @henrikdukefoss added Standard Compute ([#482](https://git

nanoMuse 1.0.5 · Lookout

🐉 nanoMuse 1.0.5 “Lookout”: the day after Compass, a second look at the whole project together with the day's reports; the daily feed is written out of sight, the Android approval card fits the screen, the desktop's dead buttons work and a quit protects the work in progress, the Windows release can update itself, and the relay's console finds anything.

1.0.4 went out on 2026-10-13. The same day we read nanoMuse through once more, this time with the reports and pull requests that arrived while Compass was being built, and nine passes again went over the relay, the Python runtime, the web console, the desktop, the Android and iPhone apps, the hosted showcase, the documentation and the workflows (#444 to #455). What they found fills this version. Some of it you will see at once: the chat the daily feed is written in no longer sits in your chat list on any device, and a routine's turn opens with its result; a long command no longer pushes the Android approval card's buttons off the screen (#436, reported by @HBSTAR0501); the ChatGPT sign-in on a phone with a proxy has a way through (#441, reported by @Kooooooo-G); the iPhone's conversation scrolls under the header, Siri understands the app's phrases in nine languages, and an emoji face set on the desktop shows on the phone (#435, reported by @Yousanflics); eight desktop buttons that quietly did nothing do their job again, and quitting the desktop asks before it ends a reply, a hands run or a task another device is waiting on. Others you will not see and should not have to: a disabled account cannot be told apart from any other through the sign-in routes, the proxy file on the desktop is the owner's alone, a hub socket that died under the desktop is noticed within a minute, and one address is one endpoint on the runtime. Beside the check, @jiangkaiqi2005 built the quit and the Windows update feed (#442, #443), @cutey-none brought DeepSeek Harness to the Python runtime (#433), and @Mola-maker's two Android fixes are in (#429, #431). Nothing moves for a person upgrading: the Android package installs over 1.0.4 and keeps everything, the desktop keeps its settings, and a 1.0.4 client works against the relay (0.26.0) unchanged.

Highlights

• The daily feed is written quietly, and a routine speaks first. On Android, the iPhone and the desktop the conversation the feed is written in no longer appears in the drawer, the chat list or search, and Write today's feed. no longer stands anywhere as a message in your voice; on Android, posts written while the app was in the background are announced by one notification that opens the Feed. The prompt a routine sends is shown as one quiet line, Routine · 〈label〉 · 08:00, and the agent is told the turn is the routine's, so it opens with the result rather than thanking you for asking (#437).
• The Android approval card fits the screen, and the ChatGPT sign-in has a way through a proxy. A long command made the card taller than the screen, so Allow, Deny and the composer sat below the bottom edge and the chat looked frozen; the card now takes at most two thirds of the screen, a command longer than six lines is folded behind Show all, and the buttons stay in view (#436). The ChatGPT sign-in sends the browser back to localhost:1455, which a proxy or VPN app on the phone can swallow; after ten seconds wi