Personal AI agent for macOS, Windows, Linux, Android & iOS. Set a goal, it works alone: coding (Skales Code), desktop + browser automation, autonomous scheduled Tasks. Teams of agents and humans across paired desktops. Agent2Agent, MCP, computer use, voice, local LLMs (Ollama), Agent Skills (SKILL.md). 26 AI providers, BYOK. No Docker, no terminal.
HTML agentic-aiai-agentai-assistantandroidautomationautonomous-agentdesktop-appelectronioslinuxllmlocal-ailocal-firstmacosmcpollamaprivacyreact-nativetypescriptwindows
SKLS: the changelog, the readme and the guide catch up with 12.7.45
Mario Simic21f3641bbf31
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex a6cae91..1e77f2c 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -6,6 +6,111 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).+## v12.7.45 - SKLS++A repair release, built the day after 12.7.4 from what the first devices+reported. The local server that the release was named for now answers on the+machines it was promised to, and a handful of screens that guessed, hid or+misnamed their failures say what actually happened instead.++### Added++- **Unsloth Desktop is a provider with its own name.** It has a tile in the+ provider grid with the same on and off switch every other provider has, an+ address already filled in, and a button that finds a running instance on+ either of the two ports it uses. It is the first local runtime that wants an+ API key, so the card says where that key is made, and a refused key is+ reported as a refused key instead of as a missing model. Refresh Models reads+ the runtime's own GGUF catalogue with the quantisation of each file, and a+ model that is still loading gets the long silent window a local runtime needs.+ How long a model stays in VRAM is set inside Unsloth Desktop, and the card+ says so rather than showing a control that would do nothing.++### Fixed++- **A turn to the local server carries a local model id.** Switching to Skales+ Local kept the previous provider's model id, so the runtime was asked for+ "openai/gpt-5-mini" it could never hold. Both request paths now resolve the+ id from the capability matrix, the provider card or the runtime itself, in+ that order, and a request to a local address can no longer carry another+ vendor's id.+- **The Mac build for Apple Silicon carries the local server.** The arm64+ package held 253 MB of engines for three other platforms and none of its+ own. Each platform and architecture now stages its own engines, a build+ gate measures the staged files against what the configs promise to ship,+ and every bundled library had its build-machine search path repaired, not+ just the executable. Mac packages shrink by the foreign payload they no+ longer carry.+- **Fetch Available Models sees what the runtime holds.** The fetch button+ knew one URL and two answer shapes while the error path knew two and three,+ so the card said "0 models" while the chat error listed all of them. One+ reader serves both now, the card offers its own library without a button+ press, and projector companion files no longer appear as answerable models.+- **Models on another disk are found.** Detection read two default folders on+ the system drive and called the rest of the computer empty. It now reads+ Ollama's configured model path including the Windows user environment, LM+ Studio's own settings, and asks running runtimes what they hold; an empty+ scan names the folders it read and offers picking one instead of claiming+ absence.+- **One status line, one Start now.** The Skales Local card rendered two+ answers to "is it running", one of them with a dead start button when no+ model was installed. One resolver decides, and the more concrete answer+ wins.+- **Connecting a Google account survives its own landing page.** The return+ page re-ran its effect the moment translations finished loading and burned+ the one-time code, which read as "This authorization is no longer open"+ seconds after a successful consent. With remote access on, the final step+ was rejected as unauthenticated and surfaced as a bare "Failed to fetch".+ The landing step is reachable on its own now, the code is claimed before+ the exchange, a run lives an hour, and a second tab can no longer erase+ the account the first one connected.+- **A tool that is named can be called.** The capability answer listed tools+ from the catalog that the running turn had not loaded, and the model,+ unable to find them, tried them as shell commands and invented an excuse+ when that failed. Naming now loads, a tool name typed as a shell command is+ steered back to the tool instead of executed, and the answer says how a+ not-yet-loaded tool becomes callable.+- **Credentials stay out of transcripts.** The integrations folder is sealed+ against file tools the way settings.json already was, and web requests+ redact tokens and secret fields by name in what they log and store.+- **Calendar reads with your account, not just your key.** With both an API+ key and a connected account, events were fetched with the key alone, which+ can only see public calendars and read as "google rejected access". A+ rejected fetch retries once with the account token, and writes always use+ credentials that can write.+- **Saved keys stop eating themselves.** Twenty fields showed a shortened+ key prefix after saving, and saving again wrote that shortening over the+ real credential, after which every check failed. A saved secret shows as+ dots that are never written back.+- **Replies to bug reports arrive.** An answer from the maintainer was+ fetched and thrown away by the notification poller, unclickable on the+ notifications page and invisible on the feedback page. It now surfaces in+ all three places.+- **Searching Settings finds every line again.** The 12.7.4 rule that let a+ section-name match win outright also erased every inventory match, which+ made "Google" unable to find the Google account setup inside Productivity+ Integrations, and the same for 144 other word-to-section pairs. Name+ matches sort first now, everything else follows under a divider, and a+ test probes every word of every section name and keyword against the index.+- **A turn that ends with nothing says what ended it.** A dropped connection,+ a filtered answer, an output limit reached before the first word, a reply+ spent entirely on reasoning and a genuinely empty answer all ended in the+ same sentence with no log line, and the retry button only existed in+ English. Each cause has its own sentence in all twelve languages, one log+ line names it on the server, and the reason and retry survive a reload.+- **Post generation stops calling a token limit a timeout.** Discover's+ compose gave every model 120 tokens; reasoning models spent them thinking+ and never wrote, and the empty result was labelled "your model did not+ answer in time". The budget starts at 1024 and climbs before reasoning is+ switched off, an answer cut at the limit is named as cut, and a fallback+ template says why it is one.+- **The diagnosis stops blaming the outside for a stop Skales ordered.** The+ server exit recorded as an external SIGKILL during Discover traffic was+ Skales's own five-second shutdown grace ending a long request; measured,+ not guessed. The forced stop marks itself and the diagnosis names it. The+ same report also claimed crash recording was off while it was on; the+ answer now comes from the installed handlers instead of an env guess.+## v12.7.4 - Skales LocalA release about screens that tell the truth. A connection that worked but readdiff --git a/README.md b/README.mdindex 76f4685..9f95a2b 100644--- a/README.md+++ b/README.md@@ -8,7 +8,7 @@ The private AI that lives on your computer and does real work, not just chat.<img width="960" height="580" alt="Dashboard" src="https://github.com/user-attachments/assets/bba3411e-85eb-44e8-ab58-be2a5a88a650" /></p><p>- <a href="https://github.com/skalesapp/skales/releases/latest"><img src="https://img.shields.io/badge/version-12.7.4-10b981?style=flat-square" alt="Version 12.7.4" /></a>+ <a href="https://github.com/skalesapp/skales/releases/latest"><img src="https://img.shields.io/badge/version-12.7.45-10b981?style=flat-square" alt="Version 12.7.45" /></a><a href="https://skales.app"><img src="https://img.shields.io/badge/license-BSL_1.1-10b981?style=flat-square" alt="License" /></a><a href="https://skales.app"><img src="https://img.shields.io/badge/Windows_+_macOS_+_Linux_+_Android_+_iOS-10b981?style=flat-square&logo=windows&logoColor=white" alt="Platform" /></a><a href="https://github.com/skalesapp/skales/stargazers"><img src="https://img.shields.io/github/stars/skalesapp/skales?style=flat-square&color=10b981" alt="Stars" /></a>@@ -255,6 +255,10 @@ Your AI agent in your pocket, live on Android and iOS. Pair via QR and the phoneNew in 12.6.0. Press the Iris button and a living particle eye ignites in its own window: a big-bang intro, then it watches, listens and speaks. No push-to-talk — it hears you when you talk, answers out loud in your native language (a choice of 55, from the first open), and stops mid-sentence when you speak over it. Tell it "morph into a car" (or a feather, or a heart) and the particles reshape into any of 1,500+ forms; while you talk, it quietly morphs along with the topic. Set a timer and the particles themselves become the countdown digits. Ask for something heavier and Iris opens the right window — Studio, Browser, Code — and tells you where it put it. Same brain and full tool set as chat, not a demo mode. Wake word "Iris" is trained on your own voice and matched locally on your machine — no audio leaves it. Right-click for new conversation, past sessions and settings. Ask her for a poem and the poem appears inside the ring, not as a caption under it — documents, search results and lists open the same way, framed by the particles themselves. On by default (Beta v1.0; switch it off under Settings → Voice and it stays off). And it is not desktop-only: Skales Mobile 2.6.0 ships the full surface natively — morphs, orbits, wake word, timer digits, barge-in, an on-surface mute and the same 55 languages with the voice following the language.+12.7.45 "SKLS" is the repair release for the one above it, built the day after from what the first devices reported. The local server the release was named for now ships for Apple Silicon as well, and answers with a local model id instead of a cloud one. The model list shows what the runtime actually holds, models on another drive are found instead of declared absent, and the Google chain tells one truth: the account connects without an error painted over the success, a tool that is named can be called, and credentials stay out of the transcript. A turn that ends with nothing now says what ended it, and a post generation that ran into a token limit stops calling it a timeout. Unsloth Desktop joins as a named provider.++12.7.4 "Skales Local" puts the inference server inside the app: no key, no account, no install. Chat, dictation, read-aloud and images can each be set to this machine first, cloud first, or off, and whoever answers is named before and after. Models are downloaded, imported or adopted from an existing Ollama or LM Studio folder, and every licence is in the app.+12.7.2 "Cockpit" is about seeing what is going on. Autopilot opens on an "In flight" list of everything running, an available update stays visible as a pill on every screen until you act on it, a tool call a model wrote as text can no longer render as the answer, WordPress grows from a posting box into running the whole site, the wait before a slow model's first token is filled with one short acknowledgment line, and mute now means mute - on every channel, at every priority, with held-back messages counted on the Notifications page instead of swallowed.12.7.1 is a same-day hotfix on top of it, from two reports that arrived within minutes of the release. GigaChat now talks to the address that actually serves its current generation, and the Russian root certificate its endpoints are signed under ships with Skales instead of being something you have to go and find. Everything that names a model now names the one that is answering: the context readout, the status pill (which is a statement about whether a chat runs locally or in a cloud), the Code window and Lio. Switching the model with a command changes that chat instead of your settings, context windows come from the provider's own catalogue, and a lost connection is reported as a lost connection rather than as a crash.diff --git a/docs/skales-guide.html b/docs/skales-guide.htmlindex e6906db..09155ea 100644--- a/docs/skales-guide.html+++ b/docs/skales-guide.html@@ -852,6 +852,7 @@<a href="#ollama-marketplace" data-chapter="providers">Ollama Model Marketplace</a><a href="#setup-koboldcpp" data-chapter="providers">KoboldCpp Setup</a><a href="#setup-lmstudio" data-chapter="providers">LM Studio Setup</a>+ <a href="#setup-unsloth" data-chapter="providers">Unsloth Desktop Setup</a><div class="section-label" data-chapter="mobile">Mobile and remote</div><a href="#ch-mobile" data-chapter="mobile">Overview and walkthrough</a>@@ -2703,6 +2704,11 @@<div class="provider-name">LM Studio</div><div class="provider-desc">Run models on your own machine with LM Studio. Point Skales at its server; no API key. Step by step below.</div></div>+ <div class="provider-item">+ <div class="provider-dot muted"></div>+ <div class="provider-name">Unsloth Desktop</div>+ <div class="provider-desc">Run models on your own machine with Unsloth Desktop, including quantisations the other local runtimes do not serve. The only local card that asks for a key, because the app issues one. Step by step below.</div>+ </div><div class="provider-item"><div class="provider-dot muted"></div><div class="provider-name">ChatGPT (Codex)</div>@@ -2932,6 +2938,25 @@<div class="tip"><strong>💡 Tip:</strong> LM Studio has a friendly UI for model management and inference parameters. Great for beginners or users who prefer a visual interface.</div></section>+ <section id="setup-unsloth">+ <h2><span class="icon">🦥</span> Unsloth Desktop Setup</h2>+ <p>Unsloth Desktop runs models on your own machine and serves them over the same OpenAI-compatible protocol Skales uses everywhere else. It is worth its own card for two reasons: it offers quantisations the other local runtimes do not, and unlike them it authenticates.</p>++ <h3>Point Skales at the server</h3>+ <p>Open <strong>Settings → AI Providers</strong> and switch the <strong>Unsloth Desktop</strong> tile on. The address is pre-filled with <code>http://127.0.0.1:8888/v1</code>, which is the port the installer uses. If your install announced a different one, press <strong>Find running Unsloth Desktop</strong>: it tries the address in the field, then 8888, then 8000, and writes back whichever answered. If none of them does, the card says so instead of spinning, and you can type the address the Unsloth window shows.</p>++ <h3>Create the key</h3>+ <p>This is the one local provider with a key field, and it is not optional. In Unsloth Desktop, open <strong>Settings → API</strong>, create a key and copy it: it starts with <code>sk-unsloth-</code> and is shown only once. Paste it into the card in Skales. If you revoke that key later, Unsloth answers every request with 401 and Skales says the key was refused and where to make a new one, rather than blaming the model.</p>++ <h3>Pick a model</h3>+ <p>Press <strong>Refresh Models</strong>. Skales reads the runtime's own catalogue, which is the model currently loaded plus the GGUF files on your disk, and keeps what each entry says about itself: its quantisation, and whether it can call tools or read images. There is no built-in list, because the only truthful one is your own machine's. Companion files such as vision projectors are filtered out, since they cannot answer a turn.</p>++ <h3>Keeping a model in VRAM</h3>+ <p>Residency is an Unsloth setting, not something a client sends per request. In Unsloth Desktop, <strong>Idle auto-unload</strong> frees VRAM after a number of idle seconds (0 keeps the model loaded, the minimum is 60) and <strong>Keep model in GPU memory</strong> holds it between prompts. If you run several models in sequence through agents or teams, that is where you tune it. Skales shows no control for it, because a switch that quietly does nothing is worse than none.</p>++ <div class="tip"><strong>💡 First call is slow:</strong> with <em>Switch model by request</em> enabled in Unsloth, a chat request naming a model that is not loaded triggers the load first. Skales allows a long silent stretch before the first token for exactly this reason, so leave it be rather than pressing stop.</div>+ </section>+<!-- ===== Chapter: Mobile and remote ===== --><section id="ch-mobile" class="chapter-lead"><div class="lead-kicker">Chapter 8</div>