Skip to content

v0.2.0-rc.1

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 05 Nov 15:51
· 33 commits to main since this release
v0.2.0-rc.1
043d9e4

Release candidate for 0.2.0. Installs of it are offered 0.2.0 when that is released.

Installers

  • Windows installer, Debian/Ubuntu and Arch/CachyOS packages, and a portable Linux archive, built by the Build installers workflow on GitHub's Windows and Linux machines. Each carries its own Node.js 22, so nothing else needs installing.
  • The LatchVerify app comes with every installer: it starts after installing (Windows), from your app menu or in the tray (Linux), and at login. Uninstalling stops it and removes its login start.
  • Installing creates your per-user latch.lock and connects only the MCP clients you choose: the Windows installer's checkbox (off by default, and updates never link anything), or latch setup afterwards on Linux (install.sh asks in your terminal). Uninstalling gives every client its original server commands back (latch setup --remove --all) and keeps latch.lock; before, only Claude Code was restored, and a client linked by hand kept pointing at the removed install.

Added

  • latch setup lists the MCP clients found and asks which to connect (--client or --all connect without asking), using the per-user lockfile: %APPDATA%\LatchVerify on Windows, ~/.config/latchverify on Linux, ~/Library/Application Support/LatchVerify on macOS. It is safe to run again, and it succeeds with a note when no client is set up yet.
  • latch setup --remove restores the client's original server entries, keeping their own settings.
  • latch setup --all and latch setup --remove --all link or unlink every MCP client on the machine at once.
  • Eight more MCP clients: VS Code Insiders, VSCodium, Cline (VS Code extension), Cline CLI, Gemini CLI, Zed, LM Studio and Amazon Q Developer, on top of Claude Desktop, Claude Code, Cursor and VS Code. Each config location was checked against the client's own docs or source. Roo Code (shut down) and Windsurf/Devin Desktop (its docs give two different paths) are left out.
  • Claude Code's per-project servers (claude mcp add in a project, kept in ~/.claude.json) are protected too. Project .mcp.json files are not rewritten: teams commit them, and a wrapped entry holds this machine's paths.
  • latch review and latch verify open the per-user lockfile when run outside a project, instead of asking for ./latch.lock. A project's own latch.lock still wins inside the project.
  • latch lock status, latch lock accept and latch lock restore check latch.lock for edits made outside LatchVerify, show what changed tool by tool, and trust or undo them.
  • Reviewers see the description a tool now serves, in latch review and in the review UI, not only that it changed.
  • LATCH_LOG_LEVEL works: debug names every frame the proxy holds, rewrites or answers itself; warn keeps stderr quiet. Warnings and errors are tagged.
  • latch watch keeps connected clients linked in the background: new servers are linked within a second, entries a client turned back are re-wrapped, entries of a moved install are repaired.
  • Desktop notifications (notify-send on Linux, a toast on Windows) when changes wait for review at start, when a proxy queues a change, when calls are blocked (repeats summed per minute) and when a new MCP client is found. Settings has switches for them, for update checks and for starting at login.
  • The MCP Servers page shows each client and its servers: which client starts which server, whether it is protected, waiting for its first run or not connected, with Connect and Disconnect.
  • The LatchVerify app (Electron): the review UI in its own window and a tray icon with the review queue count, Pause watching, Check for updates and Quit. It runs the background sync and shows native notifications; clicking one opens the review queue. Closing the window keeps it in the tray, and it starts at login unless Settings turns that off. The window loads only its own loopback server: no new windows, no navigation elsewhere, sandboxed with context isolation.
  • Opening an MCP client brings the app up. When a client starts a protected server and the installed app is not running, the proxy starts it hidden in the tray, passing it only the desktop session's variables, never the server's keys. A second start never pops the window up.
  • Updates, signed. The app checks for a new release 10 s after it starts and then once a day (Settings can turn this off), shows the release notes and installs on one click: the Windows installer, apt-get or pacman through pkexec, or the portable archive's own install.sh. Nothing is downloaded before you click. Every release publishes a latest.json signed with ed25519; the installer's sha256 and signature are checked before it runs. latch update does the same from a terminal, and latch update --check only reports. A source checkout never updates itself.
  • Dev Tools (development mode only) and pnpm demo:add add the demo servers to a client in one step.
  • Settings → Updates: the running version, the last check, Check for updates, and the release notes with Install x.y.z and restart when one is out. Works on Windows and on every Linux install (deb, Arch, portable archive).
  • Settings → Development: a switch for development mode, saved in settings.json. --dev and LATCH_DEV=1 still turn it on for one run.
  • Uninstall from the app: Settings → Uninstall LatchVerify..., the Linux menu entry's Uninstall LatchVerify action, or latch uninstall --yes. Each restores every MCP client first and keeps latch.lock. The portable install keeps its uninstall.sh, so the unpacked archive is no longer needed to remove it.
  • The Windows installer shows the app's landscape: Default Light's sunset in Windows light mode, Terminal Dark's night in dark mode, with a Welcome page.

Changed

  • Nothing is connected without your say-so. LatchVerify connects an MCP client only after you answer yes: latch setup lists the clients it found and asks, --client and --all connect without asking, the Windows installer has a checkbox, and the MCP Servers page has Connect and Disconnect. Linux packages connect nothing by themselves. Your answers are kept in settings.json.
  • Tools are pinned on a server's first run, by the proxy your client starts, instead of by LatchVerify starting each server with your keys when linking. A tool with hidden instructions is still held for review.
  • latch.lock no longer stores launch lines. Each server keeps its program name and a hash of how it is started; arguments, which can hold a key, never reach git. Older lockfiles are read and converted. latch verify takes launch lines from --config <file> (for example .mcp.json), ./.mcp.json or the client configs, and its --json report (version 2) names the program only.
  • Linked entries keep your env untouched: only command and args change, the lockfile path travels as --lockfile, and no .bak copy of the config is made any more. latch setup lists the copies earlier versions left, which hold your keys, so you can delete them.
  • The demo stays out of real installs. The Simulation page and its API (POST /api/simulate) exist only in development mode, LATCH_DEV=1. Its code and the demo tool lists load as a separate chunk only then. An installed review UI never shows sample data: when the API is down, it says so.
  • LatchVerify opens no browser. latch review --ui serves the review UI and prints its address; the app is how you open it. The latchverify-review launcher and the LatchVerify Review menu entry are replaced by LatchVerify.
  • One workflow. Pushes and merges no longer start CI. The Build installers workflow (run on demand or by a v* tag) runs the typecheck and tests on Linux and Windows first, and no installer is built or published unless they pass.
  • GET /api/info reports the running version and whether development mode is on.

Security

  • A server's first run holds back exec and credential tools, not only tools with hidden instructions, so a new server gets no more trust than a pinned one. Such a server needs one approval after it is linked.
  • A normal write can no longer sign a tampered latch.lock, and the previous version stops being trusted once a write finishes. A lockfile with no integrity record now warns, logs lock_unrecorded and never pins a first run; the app and latch watch show a notification for these and for lock_tampered.
  • Updates sign the download URL and notes, and the downloaded file's name no longer comes from the URL, which could write outside the temp folder. Releases are signed before anything is published, and fail if the key does not match the one apps ship.
  • A project's .env can no longer weaken LatchVerify: it cannot move the integrity records or the update URL, send your LLM key elsewhere, or point latch setup and latch watch at its own lockfile. The installed app does not read .env.
  • Withheld tools cannot carry an attack in their name, tool list entries with no usable name are dropped, and a tools/call that is not a well-formed request never reaches the server.
  • Your keys never leave your client configs. Env values are dropped as soon as a config is parsed and never reach LatchVerify's files, logs, API or dashboard; a canary test runs every command, the proxy and the API over configs holding a fake key and fails if it appears anywhere.
  • Guarded servers no longer get LatchVerify's own variables. The LLM API key and the paths of latch.lock and the review queue used to reach every wrapped server. No LATCH_* variable reaches a server now, and .env files only load LATCH_* lines.
  • A server cannot approve its own change by editing latch.lock. Every lockfile LatchVerify writes is recorded in an HMAC-signed integrity record under a per-user key. A proxy refuses a latch.lock the record does not vouch for, keeps the version it trusts and logs a lock_tampered event. After a git pull, run latch lock accept. This detects tampering; a program written against LatchVerify, running as the same user, can still defeat it.
  • Tier-3 tool descriptions no longer reach the model. A tool with a hidden instruction is passed on as its name and a "withheld pending review" notice, not its own text.
  • Rejections last. A rejected definition stays blocked after the proxy restarts, whatever its tier, and is not queued for review again.

Fixed

  • An install of a pre-release is offered its release: 0.2.0-rc.1 now updates to 0.2.0. Before, a release candidate never saw the release it led to.
  • The deb's in-app update no longer waits on debconf: apt-get runs non-interactively under pkexec, and a cancelled password prompt or a missing polkit agent is explained instead of reported as an exit code.
  • Connect and the app's background sync start proxies with Node.js. They used the app's own binary, so wrapped servers did not start.
  • The app keeps running when its watcher cannot start, and starts paused so the tray can retry.
  • Changing one setting no longer resets the others.
  • A proxy can no longer crash starting the app. "Start in the background" (formerly "Start at login") now also stops proxies starting it, and proxies do not try on Linux with no display.
  • Disconnecting a client saves your answer first, so a running sync cannot wrap it again.
  • Blocked calls are reported once, and the app stops polling while its window is hidden.
  • latch.lock is format version 2. Version 1 files are still read and converted.
  • A pin or approval made while latch init or latch add was running is no longer lost: their write is merged into latch.lock as it is then, under the shared state lock.
  • Each latch proxy uses about 70 MiB instead of 89, and latch watch about 75 instead of 88: commands load only what they use, and wrapped entries start node with a smaller young generation.
  • Configs with comments or trailing commas (common in VS Code's mcp.json and Zed's settings.json) are read and linked instead of skipped, and rewriting one keeps every comment: only the changed entries are edited.
  • A client config that cannot be read is reported (! cursor: ~/.cursor/mcp.json is not valid JSON: ... at line 3, column 5) instead of being skipped in silence, and it is never rewritten.
  • Config backups no longer pile up: only the 5 newest .bak copies of each config are kept.
  • npx, pnpm and other .cmd servers start on Windows, so latch init and latch setup protect them instead of skipping them.
  • Servers that print a banner on stdout can be pinned. Lines that are not JSON-RPC are skipped, and the first is reported.
  • A call to a server that was never pinned tells the model to run latch init, not latch review, which could not help.
  • The proxy never waits forever on a server: SIGTERM after 5 s without input, SIGKILL 5 s after any signal, and the server is killed if the proxy exits first.
  • Servers that ask the client something before listing their tools no longer stall: while a call waits for verification, only later calls are held, not the client's answers.
  • A single line over 64 MiB stops the server instead of growing memory without bound.
  • The queue lock is never taken over from a live holder, so a slow approval cannot lose its lock, and two waiters cannot both remove one stale lock.
  • The event log rotates at 5 MB to events.1.jsonl instead of growing forever.
  • The capture introduces itself with the real version instead of 0.0.0.

Downloads

System File Install
Windows 10/11 (x64) latchverify-0.2.0-rc.1-windows-x64-setup.exe Run it. No administrator rights needed.
Debian / Ubuntu (x64) latchverify_<version>_amd64.deb sudo apt install ./latchverify_*_amd64.deb
Arch / CachyOS (x64) latchverify-bin-<version>-1-x86_64.pkg.tar.zst sudo pacman -U latchverify-bin-*.pkg.tar.zst
Other Linux (x64) latchverify-0.2.0-rc.1-linux-x64.tar.gz Unpack, then run ./install.sh

After installing, choose which MCP clients to connect (the Windows installer asks; on Linux run latch setup). An installed LatchVerify updates itself from signed releases. See INSTALL.md.

Check a download against SHA256SUMS: sha256sum -c SHA256SUMS --ignore-missing (Linux) or Get-FileHash <file> (PowerShell).