AnyLMDownload

History

Release history

Every version built by CI, newest first. Older builds stay available, so you can roll back at any time.

v0.7.0latest

Aug 15, 2026

Release notes

v0.7.0

Styled documents across every format, research that stays on topic and finishes the job, and a smoother chat surface — with desktop installers for macOS, Windows, and Linux.

Highlights

  • Styled documents, every format. The docstyle theme kit is now ported into the app (documents/theme/): four themes (professional, academic, vibrant, informal) with palette, font pair, type scale, and margins, auto-selected from the content's tone. Applies to docx, pptx, and xlsx alike — pptx also gets a real layout engine (cover/section/content/closing slides, with card/bullet/prose/table bodies) in place of the old one-title-box-plus-one-body-box output.
  • Tool arguments get coerced, not rejected. generate_document now normalizes whatever shape a local model sends: format aliases ("Presentation" → pptx, "Word" → docx, "Excel" → xlsx), JSON-string content, and arrays of slide objects are all handled before the content guard runs. A model that used to fail four times and tell you to paste markdown into PowerPoint by hand now just produces the file.
  • Research no longer bleeds into a new chat. Fixed a Chroma recall bug where unrelated history always scored above zero and got injected as "reference" context — a fresh chat could silently inherit the previous chat's topic and even its output format. Recall now enforces a similarity floor, and naming a format explicitly ("make a PPT") pins the file type so the model can't swap it mid-turn.
  • No more silent stalls on deep research. A hard tool-round cap used to end long research turns with an empty "(stopped before any reply)" bubble. A tool-free final-answer pass now forces a written reply once the cap is hit. Repeated searches, duplicate fetches, and DuckDuckGo rate-limit responses are recognized and deflected instead of retried forever.
  • Clickable sources. Links in chat and the activity trail now render as real anchors and open in your default browser — previously any target="_blank" link, including chat markdown links, opened inside a bare, chrome-less Electron window.
  • Smoother chat surface. The composer grows with what you type instead of scrolling inside two fixed lines, and the message log now tracks new content through markdown swaps, image loads, and file cards resolving — not just snapping to bottom on append.
  • Preview generated files from any chat. Files created in a standalone (non-project) chat can now be opened in-app — PDF, Markdown, DOCX, and PPTX preview — instead of only offering "Show in folder."

Fixes

  • External links from chat, the activity trail, and OAuth callback pages now open in the system browser; the app window itself can no longer be navigated away from its own UI.
  • "(stopped before any reply)" no longer occurs after a long research turn that hit the tool-round cap.
  • Repeated web searches and re-fetches of the same URL within a turn are recognized and short-circuited instead of silently repeating.
  • "No results found." from DuckDuckGo rate-limiting is now identified as throttling (with a fallback endpoint and backoff) rather than reported as a real empty result.

What's new and internals

  • documents/theme/{tokens,detect,index}.ts, pptx-layout.ts, pptx-render.ts, docx-style.ts, xlsx-style.ts.
  • documents/normalize.ts (normalizeFormat, toMarkdown) called ahead of assertDocumentContentOrThrow in documents/index.ts.
  • memory.ts / context.ts: MIN_SCORE floors (0.45 / 0.4) on Chroma recall, reworded reference-only header.
  • intent.detectExplicit() + wantedFormat threaded through ipc.ts, agents/workers.ts, and tools/exec.ts to pin an explicitly-named format.
  • final-answer.ts (MAX_TOOL_ROUNDS, needsFinalAnswer, finalAnswerPrompt); tools/search-budget.ts (query fingerprinting, MAX_SEARCHES_PER_TURN); tools/exec.ts URL dedupe via context.fetchedUrls; web-search.ts multi-endpoint DDG fallback with backoff.
  • renderer/js/linkify.ts, rendered through rail/progress.ts and activity-trail.ts; main.ts gained setWindowOpenHandler + will-navigate handling.
  • renderer/js/autoscroll.ts (MutationObserver + ResizeObserver, follow/pin state) and composer-autogrow.ts (measured height floor/ceiling, frame-deferred first measurement).
  • project-files.ts: readGenerated / previewGeneratedFile; file-cards.ts: preview gated on file type rather than project membership, via openGeneratedViewer.

Downloads

PlatformAsset
macOS (Apple Silicon)AnyLM-0.7.0-arm64.dmg
macOS (Intel)AnyLM-0.7.0.dmg
Windows (x64)AnyLM-0.7.0-x64-setup.exe
Windows (ia32)AnyLM-0.7.0-ia32-setup.exe
Linux (x64)AnyLM-0.7.0.AppImage

Also on the site: anylm.app/download — it reads this release's assets via the GitHub API.

Upgrade notes

  • No migration required. Existing chats, projects, and settings are read as-is.
  • Document theme is auto-detected from content tone; it is not exposed as a model-facing parameter, so behavior for existing generate_document callers is unchanged unless the model previously sent malformed format/content (which now succeeds instead of failing).
  • If you lean on cross-chat memory recall, the new similarity floor means unrelated topics should no longer surface — a very sparse project may recall less than it did before.
  • Existing installs that use electron-updater will pick up latest*.yml from this release when auto-update is enabled.
  • Builds are unsigned unless Apple/Windows code-signing secrets are configured; macOS may need a one-time System Settings → Privacy & Security → Open Anyway approval on first launch.

Full Changelog: https://github.com/workvar/AnyLM/compare/v0.6.0...v0.7.0

v0.6.0

Aug 14, 2026

Release notes

v0.6.0

Knowledge specialists in the multi-agent planner, attachments that stay in context, a rebuilt PDF renderer, one rule for when tools are on, and tool approvals that survive a restart — with desktop installers for macOS, Windows, and Linux.

Highlights

  • Knowledge specialists (multi-agent Phase 2). The planner routes steps to research, fact_check, summarize, and document specialists, each with its own prompt and tool allowlist. A fact-check step is soft-inserted after research, and dependent steps receive prior step outputs instead of starting cold.
  • Per-specialist model picks. Settings assigns a model per step kind; anything unset falls back to the tool-executor model, then the chat model.
  • Attachments stay in the conversation. Images and documents are persisted as messages and re-injected on later turns, so a follow-up question can still see a file you sent earlier.
  • Much better PDFs. A dedicated print stylesheet, real Markdown table rendering, and a rewritten HTML pipeline replace the previous ad-hoc output.
  • Document turns that finish the job. After two research-only tool rounds on a turn that clearly wants a document, the model is nudged once to call generate_document.
  • One rule for "tools enabled". resolveUseTools is the single source of truth: tools default to ON, and only an explicit false turns them off, so the toggle, the project default, and the request path cannot disagree.
  • Customize is a three-step wizard. Personal context is collected in stages rather than one long form.
  • Interface refresh. Composer toolbar controls share one height, the sidebar reserves a gutter for activity dots, workspace buttons meet contrast in dark theme, and the stylesheet was reworked throughout.
  • Branded OAuth return pages. The loopback server renders success and failure pages that look like the app.
  • Unanswered tool approvals are no longer lost. A risky-tool confirmation is stored on disk and offered again next time you open that conversation — after a timeout, a stopped turn, or an app quit. Approving replays the saved call on its own, and the in-session wait grew from 2 to 10 minutes.

Fixes

  • Splash no longer hangs when a sandboxed preload crashes on launch.
  • Notification clicks reliably raise the window (focus logic moved into the window module).
  • Router no longer reads a bare "look up" as research, or a bare "document" as retrieval.
  • Async conversation patches land on the conversation they started in, even if you navigate away mid-flight.
  • Web: Open Graph image host repaired and meta copy shortened.

What's new and internals

  • agents/: Phase 2 step kinds parsed, planned, routed, and run; modelForStepKind resolves per-kind models with fallbacks.
  • documents/: new pdf-css.ts, md-html.ts, md-table.ts, auto-tools.ts, plus doc-nudge state threaded through the chat loop.
  • use-tools.ts: DEFAULT_USE_TOOLS + resolveUseTools consumed by the store, project defaults, and the renderer toggle.
  • pending-confirms.ts / pending-confirms-data.ts / confirm-resume.ts: persisted confirmations with pure, unit-tested rules and a standalone replay path.
  • oauth/pages.ts: inline-styled one-shot callback pages.
  • Renderer: conversation-target.ts, web-research-hint.ts, customize-steps.ts, restored-confirms.ts.
  • Landing page: reworked Hero and Constellation layering, globals refresh, Firebase host page updated.

Downloads

PlatformAsset
macOS (Apple Silicon)AnyLM-0.6.0-arm64.dmg
macOS (Intel)AnyLM-0.6.0.dmg
Windows (x64)AnyLM-0.6.0-x64-setup.exe
Windows (ia32)AnyLM-0.6.0-ia32-setup.exe
Linux (x64)AnyLM-0.6.0.AppImage

Also on the site: anylm.app/download — it reads this release's assets via the GitHub API.

Upgrade notes

  • No migration required. Existing chats, projects, and settings are read as-is.
  • Conversations with no stored useTools value now resolve to ON. If you relied on the old ambiguous default, check the per-chat toggle after upgrading.
  • Per-specialist model settings start empty and fall back to your chat model, so multi-agent behaviour is unchanged until you set them.
  • Existing installs that use electron-updater will pick up latest*.yml from this release when auto-update is enabled.
  • Builds are unsigned unless Apple/Windows code-signing secrets are configured; macOS may need a one-time System Settings → Privacy & Security → Open Anyway approval on first launch.

Full Changelog: https://github.com/workvar/AnyLM/compare/v0.5.1...v0.6.0

v0.5.1

Aug 9, 2026

Release notes

v0.5.1

GA4 + Clarity analytics (replacing PostHog), a first-run setup wizard, and marketing conversion events on the site — with desktop installers for macOS, Windows, and Linux.

Highlights

  • GA4 + Microsoft Clarity. Desktop product events go through GA4 Measurement Protocol in main; Clarity loads in the renderer when consent allows. PostHog is removed.
  • Same privacy model. Soft-ask consent, Privacy settings, and independent field groups still gate what leaves the device.
  • First-run setup wizard. Guided steps for bundled services, Ollama, optional embed model, and prefs before the main UI.
  • Web conversion events. Site tracks download/CTA/release views with separate NEXT_PUBLIC_* measurement IDs from the desktop app.
  • Activity strip polish. Tighter ring label layout on the marketing home page.

What's new and internals

  • Replace posthog-node with GA4 client + payload helpers; consent-gated Clarity loader; availability skip when IDs are empty.
  • Stable product taxonomy (app_opened / app_closed, auth, projects, ai_request_*, reliability) plus web download_clicked / cta_clicked / release_viewed.
  • Env swap: ANYLM_GA_MEASUREMENT_ID / ANYLM_GA_API_SECRET / ANYLM_CLARITY_ID (desktop); NEXT_PUBLIC_GA_MEASUREMENT_ID / NEXT_PUBLIC_CLARITY_ID (web).
  • Setup wizard + startup dependency checks wired into boot.
  • CI build vars updated for the new analytics secrets.

Downloads

PlatformAsset
macOS (Apple Silicon)AnyLM-0.5.1-arm64.dmg
macOS (Intel)AnyLM-0.5.1.dmg
Windows (x64)AnyLM-0.5.1-x64-setup.exe
Windows (ia32)AnyLM-0.5.1-ia32-setup.exe
Linux (x64)AnyLM-0.5.1.AppImage

Also on the site: anylm.app/download — it reads this release’s assets via the GitHub API.

Upgrade notes

  • Prefer this build over v0.5.0 if you have not installed yet — analytics stack and first-run setup differ.
  • Existing installs that use electron-updater will pick up latest*.yml from this release when auto-update is enabled.
  • Builds are unsigned unless Apple/Windows code-signing secrets are configured; macOS may need a one-time System Settings → Privacy & Security → Open Anyway approval on first launch.
  • Configure GitHub/Vercel env for the new GA4/Clarity IDs; do not put ANYLM_GA_API_SECRET in app/.env.

Full Changelog: https://github.com/workvar/AnyLM/compare/v0.5.0...v0.5.1

v0.5.0

Aug 9, 2026

Release notes

v0.5.0

Tools control that sticks, stronger document generation, privacy-first analytics, and a sharper download site — with desktop installers for macOS, Windows, and Linux.

Highlights

  • Tools toggle + defaults. Labeled Tools control in the composer, with project-wide persistence and a choice for new general chats.
  • Model lock clarity. After the first message, the model picker looks disabled and explains why on hover.
  • Research-then-write documents. Document turns prefer search → fetch → generate, with a thin-content guard so stub PDFs/DOCX/MD do not get written.
  • Privacy & analytics. Soft-ask consent on boot, Privacy settings tab, and opt-in field groups for product/reliability events (PostHog in the desktop app).
  • Site polish. Clearer download UX, product-accurate catalogs, SEO/metadata, privacy page, and brand assets.
  • Release pipeline. Split Release/Build workflows that attach macOS, Windows, and Linux installers directly to GitHub Releases.

What's new and internals

  • defaultUseTools settings + project bulk thread patch; tools scope modal for general chats; hardened toggle updates bound to the originating chat.
  • Document intent/prompt guidance, content-quality assess helper before write, and PDF CSS polish.
  • Main-process analytics client with policy filter, IPC capture surface, consent modal, and Privacy settings UI.
  • Marketing site SEO helpers, analytics, release notes markdown rendering, and download/catalog fixes.
  • CI: tag-driven release notes from docs/releases/, matrix builds, and merged latest*.yml auto-update feeds.

Downloads

PlatformAsset
macOS (Apple Silicon)AnyLM-0.5.0-arm64.dmg
macOS (Intel)AnyLM-0.5.0.dmg
Windows (x64)AnyLM-0.5.0-x64-setup.exe
Windows (ia32)AnyLM-0.5.0-ia32-setup.exe
Linux (x64)AnyLM-0.5.0.AppImage

Also on the site: anylm.app/download — it reads this release’s assets via the GitHub API.

Upgrade notes

  • Existing installs that use electron-updater will pick up latest*.yml from this release when auto-update is enabled.
  • Builds are unsigned unless Apple/Windows code-signing secrets are configured; macOS may need a one-time System Settings → Privacy & Security → Open Anyway approval on first launch.
  • Analytics is soft-ask: capture uses defaults until you Accept, Decline, or change Privacy settings. No full message bodies, tool payloads, or absolute paths are sent.

Full Changelog: https://github.com/workvar/AnyLM/compare/v0.4.0...v0.5.0

v0.4.0

Aug 8, 2026

Release notes

v0.4.0

Multi-agent orchestration, artifacts explorer, Ollama setup gate, load protection, and a full chrome restyle — with desktop installers for macOS, Windows, and Linux.

Highlights

  • Multi-agent core. Orchestrator plans and runs parallel worker steps with settings, token metering, and a collapsed agent trail in chat.
  • Artifacts explorer. Browse, open, reveal, and delete generated files. Standalone chats can write documents to Documents/AnyLM/generated.
  • Ollama setup gate + boot splash. Clearer first-run path when Ollama is missing or not running.
  • Load protection. Soft-stop turns when system RAM crosses the kill threshold, with a settings toggle for the cutoff.
  • Project-first coding. Coding turns prefer a real project folder scaffold path.
  • Web research. Skill + URL chip to enable research per chat or globally, with stronger follow-up prompts.
  • Reasoning stream. Thinking/reasoning tokens surface live with a global Working strip.
  • Models browser. Live catalog with list/grid and RAM compatibility cues.
  • Settings hub. Left-nav sections, Agents panel, and tabbed project settings.
  • Site-theme chrome. Constellation brand, glass surfaces, auth and sidebar restyle.

What's new and internals

  • Multi-agent domain types, planner/classify wrappers, scheduler, and orchestrator wired into chat:start.
  • Standalone document destination + Artifacts sidebar explorer over generated and project folders.
  • Boot splash / Ollama detect-start gate; load-protection soft-stop in the turn pipeline.
  • Web research skill recovery, conversation skillOverrides, and URL-chip hint UI.
  • Reasoning token streaming and global Working strip selectors.
  • Electron site-theme redesign across auth, sidebar, and main chrome.

Downloads

PlatformAsset
macOS (Apple Silicon)AnyLM-0.4.0-arm64.dmg
macOS (Intel)AnyLM-0.4.0.dmg
Windows (x64)AnyLM-0.4.0-x64-setup.exe
Windows (ia32)AnyLM-0.4.0-ia32-setup.exe
Linux (x64)AnyLM-0.4.0.AppImage

Also on the site: anylm.app/download — it reads this release’s assets via the GitHub API.

Upgrade notes

  • Existing installs that use electron-updater will pick up latest*.yml from this release when auto-update is enabled.
  • Builds are unsigned unless Apple/Windows code-signing secrets are configured; macOS may need a one-time System Settings → Privacy & Security → Open Anyway approval on first launch.

Full Changelog: https://github.com/workvar/AnyLM/compare/v0.3.0...v0.4.0

v0.3.0

Jul 6, 2026

Release notes