Repository navigation
Releases: mudrii/compme
Release list
v0.1.7
Full Changelog: v0.1.6...v0.1.7
v0.1.6
v0.1.5
v0.1.4
v0.1.2
Compme v0.1.2 — model loading fixed + model management
Fixes the main v0.1.1 problem — a downloaded model never actually loaded — and
adds two ways to manage models. All inference stays local; no telemetry
(CI-enforced).
Fixes
- Downloaded models now load. v0.1.1 only wired a model on a fresh
download's completion, so models already on disk (downloaded under an older
build) were never loaded — the Setup "Model file" row stayed ✗ and re-clicking
Download reported "already present" without wiring anything. On launch Compme
now auto-adopts the newest.ggufin its models folder and persists
COMPME_MODEL_PATH; the Download button wires an already-present model too.
New
- Show Models Folder — a Setup-tab button that opens
~/Library/Application Support/compme/modelsin Finder. - Choose Model… (bring your own model) — pick any local
.gguffrom a file
panel; it's validated (GGUF header) and used in place, no copy or re-download. - New app icon — an inline-completion motif (typed text · caret · ghost
suggestion).
Upgrade
brew upgrade --cask compmeReopen Compme. If a model was downloaded but never loaded, it's adopted
automatically on launch (one relaunch may be needed the first time).
Still ad-hoc signed (no Apple Developer ID yet): Gatekeeper blocks the first
launch — approve under System Settings → Privacy & Security ("Open Anyway"), or
install the cask with --no-quarantine. Requirements unchanged (macOS 14+
Apple silicon, Accessibility permission).
Full Changelog: v0.1.1...v0.1.2
v0.1.1
Compme v0.1.1 — model setup + app icon fixes
A small follow-up to v0.1.0 fixing two problems people hit on the first
Homebrew install: a downloaded model never becoming usable, and a missing app
icon. All inference stays local; no telemetry (CI-enforced).
Fixes
- Downloaded models are now wired up automatically. Previously a completed
download only logged a hint to setCOMPME_MODEL_PATH— impossible for a
Finder-launched.app— so the Setup "Model file" row stayed unchecked and
the model was never loaded. The app now persists the downloaded model path to
its config; relaunch and it loads. - Setup pane shows download progress and errors. Download status (percent,
"downloaded — relaunch to use", or the failure reason) was previously
stderr-only and invisible in a GUI launch; it now appears in the Setup tab. - App icon. The bundle now ships an icon (
CFBundleIconFile), so Finder and
the Dock no longer show the generic placeholder.
Upgrade
brew upgrade --cask compmeThen reopen Compme. If you downloaded a model in v0.1.0, click Download
Model once more (or relaunch after it finishes) so the model path is recorded.
Still ad-hoc signed (no Apple Developer ID yet): Gatekeeper blocks the first
launch — approve under System Settings → Privacy & Security ("Open Anyway"), or
install the cask with --no-quarantine. Requirements are unchanged from v0.1.0
(macOS 14+ Apple silicon, Accessibility permission, ~1 GB for the default
model).
Full Changelog: v0.1.0...v0.1.1
v0.1.0
Compme v0.1.0 — first public release
Compme is an open-source inline text-completion engine for macOS: local GGUF
completions rendered as ghost text in the field you are typing in, accepted
word-by-word or in full from the keyboard. All inference is local
(llama.cpp/Metal); there is no telemetry of any kind — enforced by a CI policy
gate, not just policy.
Highlights
- Inline completions — short local-model continuations, debounced and
gated per field, with candidate cycling and a force-show hotkey. - Rebindable accept keys — Tab = next word,
`= full accept by
default; any key plus Shift/Ctrl/Option/Command via the in-app key recorder. - Writing aids — high-precision trailing-word autocorrect, standalone
grammar/spell fix (underline + banner + exact range replacement),:shortcode
emoji with skin-tone/gender preferences, curated thesaurus, opt-in US→UK
spelling normalization. - Per-app and per-domain control — enable/disable, Tab-key and collection
opt-outs, and feature overrides per app; browser-domain exclusions fail
closed when no fresh URL resolves. - Privacy by construction — prompts pass a redaction layer
(secrets/emails/cards) before any model call; optional typing memory is
AES-256-GCM encrypted with a Keychain-managed key and redacted before
encryption; screen-context OCR is redacted at the sole capture seam. - Model picker — built-in catalog with per-machine RAM-fit verdicts and
verified, resumable downloads (SHA-256 checked before use). - Settings — nine-tab settings window, menu-bar tray, launch-at-login,
signedcompme://deep-link overrides (Ed25519, fail-closed).
Requirements
- macOS 14+ on Apple silicon (Metal inference).
- Accessibility permission (prompted on first run). Input Monitoring is not
required. - ~1 GB disk for the default small model (downloaded on first setup).
Install
Homebrew (available once this release is published):
brew tap mudrii/compme https://github.com/mudrii/compme
brew install --cask compmeOr download compme-0.1.0-macos.zip below; verify with the .sha256 file.
compme-0.1.0-update.json is a machine-readable manifest for tooling (nothing
in-app consumes it yet).
This release is ad-hoc signed (no Apple Developer ID yet): Gatekeeper
blocks the first launch. After installing, approve the app under System
Settings → Privacy & Security ("Open Anyway"), or install the cask with
--no-quarantine if you accept the trade-off. A future release will be
Developer-ID signed and notarized.
Known limitations
- macOS only. Windows and Linux adapters are committed deliverables: the
sharedPlatformAdaptercontract, fail-closed scaffolds, Windows host
services (file hardening, console handling), and 3-OS CI already ship; the
real UIA/AT-SPI adapters are the next roadmap tier. - Inference is Metal-only in this build (CPU fallback builds exist for other
targets but are not packaged). - Secure fields and secure-input sessions are always blocked (by design).
Full Changelog: https://github.com/mudrii/compme/commits/v0.1.0