Skip to content

Releases

v2.5.0 — easier to reply, easier to start

Choose a tag to compare

@github-actions github-actions released this 10 Nov 06:58

v2.5.0 — easier to reply, easier to start

Big usability pass. Nothing new to connect, a lot less to trip over.

Replying

  • One Reply window. It shows who you're writing to and on which service, and it sends exactly the text you see.
  • Every send still waits 5 seconds so you can undo it. Once it's too late, Undo goes away instead of pretending.
  • The Ask panel only answers questions now. It can't send anything.
  • A send that fails stays in the Act queue under "Couldn't send" with the reason and "Try again". That includes a service that isn't connected.

First launch

  • You can start with sample messages and try the whole app before granting any access. Nothing is sent while you're in the sample. The sample now has a WhatsApp chat too.
  • The AI step no longer blocks you on an API key, and there's help for Full Disk Access.

Easier to read

  • Plainer labels in digests and the Act queue: "3 messages · 1 person", "Uses saved notes", "Show 3 source messages", "Optional".
  • The AI draft button only shows up when an AI provider is set up.

Settings

  • Service switches save right away, and unsaved fields aren't lost when you close the window.
  • The WhatsApp card tells you exactly what's missing.
  • MCP asks you to confirm before turning on, shows setup steps for Claude and Cursor, and shows when an agent last asked for something.

Privacy

  • Fixes the two MCP issues listed under "Known issues" in v2.4.0, plus one more: digests that draw on Local-only conversations are withheld from MCP, and an outside agent can't change a reply you already reviewed.
  • LLMessenger no longer asks for microphone or speech recognition access.

Accessibility

  • Better VoiceOver announcements, visible keyboard focus, larger small text, stronger contrast.

Good to know

  • After upgrading, the first few digests per service aren't available over MCP until new ones are generated.
  • Still unsigned: right-click → Open on first launch.

v2.4.0

Choose a tag to compare

@github-actions github-actions released this 09 Nov 21:47

Two new opt-in features: WhatsApp (read-only) and an MCP server mode. Both are off until you turn them on.

WhatsApp, read-only

  • LLMessenger can now read WhatsApp Desktop's local history on your Mac, the same way it reads iMessage, so WhatsApp chats show up in digests and owed replies.
  • There is no local way to send through WhatsApp, so LLMessenger never sends there. It drafts the reply; you copy it and answer in WhatsApp.
  • Enable it in Settings → Services. Requires WhatsApp Desktop installed and signed in, and Full Disk Access.

MCP server mode

  • LLMessenger --mcp lets an MCP client you configure (Claude Desktop, Claude Code, Cursor) read your owed replies and latest digest, list conversations, and search your messages, all locally over stdio.
  • An agent can only propose a reply. It lands in your Act queue labelled as coming from an external agent, is never part of a batch send, and is sent only if you approve it.
  • Conversations marked Local-only are never exposed, and are left out of digests while MCP is on. Global Local-only mode blocks all MCP reads.
  • Enable it in Settings → AI & Privacy → Privacy, where the config snippet for your client is shown.

Also

  • If the database fails to open, the app now explains and asks before moving it aside, instead of silently starting fresh. Automatic recovery still happens for a genuinely corrupt file.

Known issues, fixed on main and coming in the next release

These only matter if you turn MCP on:

  • A connected agent could replace the text of its own pending proposal after you had read it.
  • A digest could still be served over MCP after one of the conversations it was built from was marked Local-only.

Install

The app is unsigned: after unzipping, right-click LLMessenger.app → Open → Open on first launch (or System Settings → Privacy & Security → Open Anyway). Verify the download against SHA256SUMS.

Full Changelog: v2.3.3...v2.4.0

v2.3.3

Choose a tag to compare

@github-actions github-actions released this 28 Sep 13:18

A 5-agent Opus review of the v2.3.2 bug-hunt session (correctness, SwiftUI lifecycle, security, test quality, docs/release consistency) found real problems in what that session shipped. This release fixes the ones worth a new binary for.

Fixed

  • Refresh no longer clears a question you were mid-typing. v2.3.2's fix for draft leakage across digest navigation was too broad — it also fired when the app auto-selected a new digest after a refresh finished, wiping what you were writing. Only navigating to a different digest yourself now clears the composer.
  • An abandoned @-mention target could silently redirect a reply to the wrong person. Picking someone from the @ picker, then abandoning it and pressing REPLY on an unrelated owed reply, could send to the abandoned target instead.
  • A priority rule that suppresses every message now asks for confirmation before saving. A rule with no contact, keyword, or service condition — just "suppress" — matches every message, including urgent ones, with no warning until now.
  • Closed a narrow window where auto-send could be saved without the confirmation dialog being shown. The v2.3.2 fix for a spurious confirmation popup relied on scheduler timing that wasn't actually guaranteed.
  • The About tab now reports the correct version and license. The v2.3.2 binary said it was v2.3.1 (Info.plist was never bumped) and still said "MIT License" instead of Apache 2.0.
  • The accessibility-audit fix from v2.3.2 was a no-op, replaced with one that actually works — a SwiftUI .accessibilityHidden modifier can't affect an AppKit-owned window's accessibility state.
  • Removed the Homebrew install instructions from the README — the tap has no LLMessenger cask yet, so the command failed for anyone who tried it.

Verification

730 tests green (6 new since v2.3.2), zero static-analyzer warnings, red→green verified for every regression test added.

Full Changelog: v2.3.2...v2.3.3

v2.3.2

Choose a tag to compare

@github-actions github-actions released this 26 Sep 06:11

Full Changelog: v2.3.1...v2.3.2

v2.3.1 - Activity becomes a trend

Choose a tag to compare

@googlarz googlarz released this 14 Jul 14:38

Small, focused release: Activity now shows a real 14-day trend of when things needed you, instead of just today.

What changed

  • Activity covers 14 days, grouped by day, instead of a single "today" list.
  • Today expands automatically; older days collapse to a header with a "N needed you" count — tap to expand. Two weeks of history stays scannable instead of becoming a wall of rows.
  • Quiet days show no badge. The count only appears when something actually needed you, so a clear stretch reads as calm.
  • Fixed the "old ledger" look in Activity and per-conversation history — entries now separate by whitespace with one clear headline per row, not a hairline after every line.

Deliberately not a calendar heatmap — a quiet mono count per day header does the job without adding a chart widget to an otherwise typographic app.

Install: brew tap googlarz/tap && brew install --cask llmessenger

Verified: 717 tests pass.

🤖 Generated with Claude Code

v2.3.0 - The Mac-native redesign

Choose a tag to compare

@googlarz googlarz released this 13 Jul 19:38

The largest UI release since 2.0 — LLMessenger now behaves like a first-class macOS citizen.

Navigation, rebuilt

  • A proper sidebar (Act · Digests · Activity) where every selection predictably drives the main content — the fix for tabs and detail content being unrelated.
  • Act and Activity are full-width workspaces; Digests pairs the archive list with the open digest, Mail-style.
  • Sidebar visibility, selected section, and window state persist across launches. Narrow windows collapse from three columns to two — nothing squeezes or clips down to 860×520.

Native chrome

  • Real title-bar toolbar: sidebar toggle, contextual digest back/forward (⌘[ / ⌘]), window title per section, native search field (⌘F), refresh, service status, and a More menu.
  • A full macOS menu bar — App/File/Edit/View/Window/Help with every command and shortcut (the app previously had none).

Calmer reading

  • Digest cards start compact — headline, two summary lines, one action — and expand on demand. Only the most urgent card opens expanded.
  • Reading is reading: the AI composer is now an Ask panel opened from the toolbar; tone controls appear only while writing; the AI provenance line lives in the status popover instead of a permanent footer.

One work queue

  • Act is a single ranked queue: Needs your decision → Ready to send → Waiting on others → Later.
  • Safety information appears where it matters — on the send countdown and batch bar — not as a permanent strip. Batch approval appears only after ⌘-selecting compatible sends.

System-wide polish

  • Settings consolidated to five panes (AI & Privacy, Services, Behavior, Schedule, About).
  • Reduce Motion and Increase Contrast honored everywhere; no text below 10pt.
  • Refresh is a toolbar spinner; confirmations are floating toasts that never displace content.

Install via Homebrew: brew tap googlarz/tap && brew install --cask llmessenger

Verified: 717 tests, 0 failures, including a live-window accessibility audit.

🤖 Generated with Claude Code

v2.2.7 - Cleaner internals, faster Desk, in-app release notes

Choose a tag to compare

@github-actions github-actions released this 02 Jul 04:52

v2.2.7 is a hardening and performance release: no new concepts, just a codebase that is safer to run, snappier to use, and easier to review.

Stability

  • Removed five runtime force-unwraps that could crash the app on unexpected state: message grouping in the Signal and iMessage adapters, the iMessage poll fallback, demo seeding, and memory compression.
  • The two agent safety switches (whole-engine off vs. delegation off) are now named constants with documentation making clear they must never be merged.

Performance

  • Brief JSON is decoded once and cached (validated by content hash) instead of re-decoded on every render — the Today summary, first-digest banner, chat header, outcome stats, weekly recap, and menu-bar previews all shared this cost.
  • The Act feed builds and sorts its item list once per render pass instead of ~4 times, and no longer risks a synchronous database read from inside a view update.
  • Menu-bar updates coalesce into one rebuild per refresh cycle instead of 25+.
  • Per-row date formatting now uses shared cached formatters; the needs-reply query moved off the main thread.
  • Verified behavior-identical: the Act surface renders pixel-identical (0 of 912,000 pixels changed) to the previous release.

Maintainability

  • AppState (1,336 lines, one file) split into 8 focused extension files at its existing section boundaries — code moved verbatim, with the gated auto-send section explicitly marked as the sacred path.
  • Shared ConversationAccumulator replaces duplicated conversation-grouping code in the Signal and iMessage adapters.
  • Service-health reads consolidated onto a single live store.

In-app release notes

  • The About tab now has a "What's new" section showing recent version highlights — no GitHub visit needed. Notes ship in the same commit as each version bump.

Validation

  • Full test suite: 620 tests, 1 skipped, 0 failures — including the complete delegation and injection suites run against the relocated auto-send code.
  • Design snapshot suite green; About-tab snapshot added as a new baseline.

Full Changelog: v2.2.6...v2.2.7

v2.2.6 - First-run trust and recovery polish

Choose a tag to compare

@googlarz googlarz released this 01 Jul 15:32

v2.2.6 is a trust, first-run, and recovery polish release. It focuses on making the product easier to understand on day one, safer to correct, and clearer about why each brief/draft exists.

First-run confidence

  • Added a live setup diagnosis for AI, services, messages, and privacy while the first real digest is loading.
  • Demo mode is now positioned as the safe command center to explore while real messages, contacts, and context sync.
  • The first real digest moment now includes a short guided checklist: open sources, mark one done, draft only if ready, quiet noise.
  • If no LLM provider was selected yet, the app now falls back to a local-first provider path instead of starting from an unconfigured dead end.

Trust and explainability

  • Brief cards now show plain confidence labels such as High confidence, Context assisted, or Check sources.
  • The “Why this card” section now includes an explicit confidence explanation alongside reason, evidence, memory, and privacy.
  • Reply drafts now carry provenance so users can see why a draft was created and what context shaped it.
  • Local-only and never-draft conversation privacy now blocks draft creation in more paths when a cloud client would be involved.

Recovery everywhere

  • Added visible undo receipts for skipped suggested actions, dismissed owed replies, snoozed owed replies, card completion, archive/snooze flows, and conversation-context edits.
  • Added “Not a reply” learning feedback so users can correct false reply-needed cards quickly.
  • Product-health and learning surfaces now count drafts and undo use, making the app’s review-first behavior more visible.

Product and docs polish

  • Desk tabs now include plain-language context: Act is what needs you now, Digest is what happened, Activity is what changed.
  • PRD now has a current v2 status note so stale v1 roadmap items do not mislead future planning.
  • Refreshed design snapshots for the updated first-run and desk surfaces.

Validation

  • Full local test suite: 619 tests, 1 skipped, 0 failures.
  • Design snapshot suite: 8 tests, 0 failures.
  • Final Debug build passed after aligning app and widget versions to 2.2.6.

Full Changelog: v2.2.5...v2.2.6

v2.2.5 - Act tab focus, first-week guide dismiss

Choose a tag to compare

@github-actions github-actions released this 01 Jul 05:39

Fix-forward polish on v2.2.4. No new features — this corrects where the v2.2.4 "product-love" surfaces live so the Act tab goes back to being the lean, action-first primary view.

Act tab: back to "what should I do now?"

  • v2.2.4 put TodaySummaryView, FirstWeekGuideView, OutcomeStripView, and LearningReceiptsView in DeskView's shared header — above the tab bar, so they rendered on every tab (Act, Digest, Activity), not just where they made sense. That buried the actual action queue under up to four stacked summary strips.
  • TodaySummaryView and FirstWeekGuideView now live at the top of the Act tab only (ActFeedView) — both are queue/action/promise oriented ("people waiting", "actions ready", teach-the-primary-loop).
  • OutcomeStripView and LearningReceiptsView moved to the Activity tab, alongside the existing WeeklyRecapView / ProductHealthView dashboard — both are retrospective ("what got done", "what did it learn").
  • DeskView's shared header is back to just the demo-transition/kill-switch banner + ToDoStripView (genuinely cross-tab: open commitments, tasks, the Maybe bucket).

First-week guide: real dismiss

  • shouldShowFirstWeekGuide()'s OR-gate (openedDigests==0 || handledCards==0 || !hasLearningSignal || activeDays<2 || suggestionCount>0) could resurrect the guide mid-week even after a user closed it — none of those conditions tracked an explicit dismiss.
  • Added ProductLoveMetrics.guideDismissed, persisted the same way as the other UserDefaults-backed counters, plus a dismiss × button on the guide's header row. A dismiss now wins over the OR-gate permanently.

Validation

  • 615 local tests, 1 skipped, 0 failures (adds one regression test: dismissing the guide holds even when a fresh suggestion would otherwise reopen it).
  • Verified visually via the offscreen DesignSnapshotTests harness (testSnapshotAct, testSnapshotEmptyState) before tagging — confirmed the render order top-to-bottom: To Do → Maybe → tab bar → Today summary → First Week guide (with visible dismiss ×) → action feed.

Full Changelog: v2.2.4...v2.2.5

v2.2.4 - Trust, demo, and product love

Choose a tag to compare

@github-actions github-actions released this 01 Jul 05:21

What's new

Try before connecting accounts

  • Onboarding now recommends Open demo first, so new users can explore the real command center with synthetic conversations before granting message access.
  • The empty state now offers a direct sample command-center CTA instead of leaving users at a setup-only screen.

Stronger privacy controls

  • Card context popovers now include per-conversation privacy controls: Normal, Local-only, and Never draft.
  • Privacy override changes preserve existing conversation labels, priority hints, and learned context instead of wiping surrounding state.

More trust in every brief

  • Cards now expose clearer trust cues: source count, reason, context/memory use, and privacy posture.
  • A new source-evidence screenshot shows how card claims map back to local stored messages.

Product-love surfaces

  • Added product-health, first-week, outcome, trust, learning, queue, and recap surfaces to make progress and confidence visible.
  • One-click VIP/quiet teaching is easier to discover from cards.

README and screenshots

  • README now has a Try demo first flow with demo, Act queue, and source-proof screenshots.
  • README FAQ explicitly states screenshots use synthetic demo fixtures, not private inbox data.
  • Added regenerated synthetic screenshots: demo mode and source evidence.

Validation

  • GitHub Actions release build succeeded on macOS.
  • Local full test suite passed before tagging: 614 tests, 1 skipped, 0 failures.
  • Release assets attached: LLMessenger-unsigned.zip and SHA256SUMS.

Full changelog: v2.2.3...v2.2.4