Open source Grok Bot alternative, Bring your own agent and subscription: message Claude Code, Codex, Gemini CLI, Cursor, Copilot and 40+ AI agents as bots on your own computer, from iPhone, Mac, Windows, Linux or terminal. Agent-agnostic and free. Also a Muse and Dots alternative. Built in Rust.
Windows testing
Codync ships x64 and ARM64 Windows apps (Electron + the matching bundled codync-host.exe). Real Windows behavior is
checked in a UTM virtual machine on the Mac. Everyday end-to-end checks stay on the Mac (minimal
e2e); use the VM for a full e2e, or when a change or a bug depends on Windows itself (paths,
cmd.exe, .cmd shims, the service, timing). The VM slows the whole Mac down: start it for the
check and stop it as soon as the check is done.
Host CI runs cargo nextest run --workspace --profile ci on every OS (host/.config/nextest.toml):
each test runs in its own process, so fixture hosts no longer contend on process-global memory scan
locks and Windows runs them in parallel too. A test that fails and then passes on a retry is
reported as FLAKY instead of failing the run: fix it rather than lean on the retry.
Desktop release CI builds x64 on windows-latest and ARM64 on windows-11-arm, then smoke-tests
each installed host after running the installer and checking the native app/host executable headers.
Installers, checksums and blockmaps use codync-windows-<arch>.exe names.
Both builds contribute to one latest.yml update feed, uploaded only after both installers;
electron-updater selects the file matching the installed app's architecture. The pinned computer-use
driver and UI Automation helper remain x64 and run through Prism on ARM64 Windows 11, as do
registry agents that ship only a windows-x86_64 build (the ARM64 host downloads that one instead).
Windows packaging forces the supported BCJ archive filter: the bundled NSIS decoder silently skips
executables compressed with modern 7z ARM64 filters, even when the installer reports success.
For a manual CI check, dispatch Release desktop with windows_only enabled to build both
Windows architectures without requiring Mac signing secrets or running Linux packaging.
The VM
- UTM "Codync Windows": Windows 11 ARM64 (use the ARM64 build for native checks, x64 for Prism emulation),
8 GB / 4 cores / 64 GB NVMe, account
codync/codync, signs in automatically. - Start / stop from the Mac, never through UTM's window:
osascript -e 'tell application "UTM" to start virtual machine named "Codync Windows"'(stop … by forceonly when a normal stop hangs;query ip of …works).utmctlfails with OSStatus -10004. Pressing UTM's Run button through accessibility crashed UTM once; if UTM is in a bad state, quit it and reopen it in the background withopen -g -a UTM. After stopping, end leftover sessions withpkill -f "ssh codync-win". - Shell:
ssh codync-win(PowerShell). The~/.ssh/configalias points at192.168.64.4with the key~/.ssh/codync_windows_vm. SSH is up a minute or two after start:until ssh -o ConnectTimeout=5 -o BatchMode=yes codync-win 'echo up'; do sleep 10; done. - GUI apps must start in the signed-in session (session 2):
schtasks /create /tn X /tr <exe> /sc once /st 23:59 /it /f; schtasks /run /tn X. - Screen without touching the Mac's UI: QEMU VNC on
127.0.0.1:5905, e.g.uvx --from vncdotool vncdo --force-caps -s 127.0.0.1::5905 capture x.png. Input lags a few seconds;--force-capsis needed for symbols. - Network: the VM's internet is slow and drops connections (2026-10-20: a 4.4 MB npm tarball
took 27 s in the VM and 0.3 s on the Mac, about 165 KB/s; a 130 MB
Invoke-WebRequestdownload stopped at 45 MB without an error). Download large files on the Mac andscpthem in. - The Mac sleeping pauses the VM: wrap long installs in
caffeinate -i. - Rebuilding the VM: the unattended install assets live in
~/VMs/codync-windows/(about 20 minutes). NVMe must be the first drive and both ISOs detached afterwards, or Windows reinstalls on every boot.
Installing a release
Download the installer on the Mac and copy it in under a new name (
scpcan't overwrite a file Windows still holds open:dest open … Failure):gh release download v2.20.3 --repo leepokai/Codync -p codync-windows-x64.exe -p codync-windows-x64.exe.sha256 -D <dir> shasum -a 256 <dir>/codync-windows-x64.exe # must match the .sha256 file scp <dir>/codync-windows-x64.exe codync-win:codync-2.20.3.exeStop Codync, then install silently over SSH:
Get-Process Codync,codync-host,codync-hostw -ErrorAction SilentlyContinue | Stop-Process -Force $i = Start-Process "$env:USERPROFILE\codync-2.20.3.exe" -ArgumentList "/S","/currentuser" -Wait -PassThru; $i.ExitCode & "$env:LOCALAPPDATA\Programs\codync-desktop\resources\codync-host.exe" --versionExit code 2 is the installer's integrity check: the file is truncated, so compare
(Get-FileHash <file>).Hashwith the release's.sha256.taskkillcan leave the old binaries locked;Stop-Process -Forceon all three processes frees them.Start the app in the signed-in session (
schtasksabove, with$env:LOCALAPPDATA\Programs\codync-desktop\Codync.exe). The first launch asks Defender Firewall aboutcodync-host; the host then starts at sign-in through the HKCU Run valueCodyncHost(codync-hostw.exe).Turn analytics off on the test host before anything else, so the VM never counts as a user:
setAnalytics {"enabled": false}.helloreports the current setting (analytics).
Driving the host over SSH
The API is loopback-only, which SSH satisfies. Copy this helper to the VM as api.ps1:
param([string]$Method, [string]$Body = "{}")
# A body of "@file.json" is read from that file (avoids quoting JSON through ssh).
if ($Body.StartsWith("@")) { $Body = Get-Content -Raw $Body.Substring(1) }
$t = Get-Content "$env:USERPROFILE\.codync\token"
$r = Invoke-RestMethod -Method Post -Uri "http://127.0.0.1:19222/api/$Method" -Headers @{ Authorization = "Bearer $t" } -ContentType "application/json; charset=utf-8" -Body ([System.Text.Encoding]::UTF8.GetBytes($Body))
$r | ConvertTo-Json -Depth 12
- Run it as
powershell -NoProfile -ExecutionPolicy Bypass -File api.ps1 <method> "@<body>.json": Windows refuses scripts by default, and JSON quoted through ssh and PowerShell arrives broken, so write each body to a file on the Mac andscpit over. refreshBackendslists agents (installed,available,signedIn);agentAuth({"backend":"claude"}) runs the sign-in check the agent sheet shows;createBot({"name":…,"backend":"claude","cwd":"C:\\Users\\codync","connectors":[],"skills":[]}),send({"botId":…,"text":…}) andhistory({"botId":…,"limit":20}) cover most agent checks. A failed start or a sign-in card shows up inhistoryas a notice (data.text,data.connectionRequest).- Waiting for a reply: poll
historyevery few seconds until an entry that isn'tuserappears. Agent starts in the VM take minutes, so run the wait as a background command on the Mac rather than one long foregroundssh. - Host log:
%USERPROFILE%\.codync\host.log(strip colors with-replace "\x1b\[[0-9;]*m",""). Agent processes:Get-CimInstance Win32_Process | Where-Object { $_.Name -like "node*" } | Select ProcessId,ParentProcessId,Name,CommandLine.
Agents
Install in the agent sheet runs the CLI's native installer. Agents whose ACP adapter comes from npm (Claude, Codex, Gemini, Qwen, …) need Node.js: without one on the VM, Extensions shows Add Node.js, which puts Codync's own copy in
%USERPROFILE%\.codync\node(see agent setup). To test the computer's own Node instead,winget install OpenJS.NodeJS(the VM has Node 24, ARM64).Signing an agent in needs the user: it's an interactive browser sign-in on the VM's screen, with their account. Ask them; never sign in on their behalf. Claude Code in the VM is signed out.
First npm adapter download: Claude's adapter with its Claude binary is about 340 MB, and on the VM's network the host's 20-minute prefetch timed out (
couldn't download Claude Agent with npm: timed out). Seed the npx cache from the Mac instead. The folder name is the first 16 hex digits of the SHA-512 of the package spec;--cpumust match the VM's Node (node -p process.arch), not Codync's x64:P=@agentclientprotocol/[email protected] printf '%s' "$P" | shasum -a 512 | cut -c1-16 # npx cache key, e.g. 1f872328d1400346 mkdir -p <dir> && printf '%s\n' '{"dependencies":{"@agentclientprotocol/claude-agent-acp":"^0.94.0"},"_npx":{"packages":["@agentclientprotocol/[email protected]"]}}' > <dir>/package.json npm install --prefix <dir> --os=win32 --cpu=arm64 --no-audit --no-fund COPYFILE_DISABLE=1 tar -C <dir> -czf npx-claude-win.tgz . && scp npx-claude-win.tgz codync-win:$d = "$env:LOCALAPPDATA\npm-cache\_npx\<key>" New-Item -ItemType Directory -Force $d | Out-Null; tar -xzf "$env:USERPROFILE\npx-claude-win.tgz" -C $d Remove-Item -Recurse -Force "$d\node_modules\.bin"; cd $d; npm rebuild --offline --no-audit --no-fundThe Mac's
node_modules\.binhas symlinks, not Windows.cmdshims;npm rebuildwrites them (check forclaude-agent-acp.cmd).Launches stay unreliable even with a seeded cache:
npx -yasks the registry on every launch and fetches anything missing (one 2026-10-20 launch spent 55 s on a single small tarball), so an adapter can sit silent past the host's start limit (the agent didn't start within 180s), with or without--prefer-offline. When that blocks a check, run the agent's commands by hand (next section) or reproduce the condition on the Mac.
Checking an agent by hand
Time a CLI the way the host or adapter runs it, e.g.
foreach ($i in 1..3) { $s = Get-Date; & "$env:USERPROFILE\.local\bin\claude.exe" auth status | Out-Null; ((Get-Date) - $s).TotalSeconds }(2026-10-20: 5.5 to 8.8 s in the VM).See what an ACP adapter reports: a small Node script that starts the adapter's command with
shell: true(for a cached npm adapter,…\_npx\<key>\node_modules\.bin\claude-agent-acp.cmd), writes oneinitializeline, and prints every line it gets back with a timestamp for 25 s (stderr passes through):{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":1,"clientCapabilities":{"fs":{"readTextFile":false,"writeTextFile":false},"terminal":false,"auth":{"terminal":true},"_meta":{"terminal-auth":true}}}}An adapter that advertises
agentCapabilities._meta.authStatusshould follow its answer with an_auth/status_updatenotification.npm: run the host's command yourself with
--loglevel http(orsilly) and-RedirectStandardError; npm's own logs are in%LOCALAPPDATA%\npm-cache\_logs.
Testing an unreleased host
Cross-build the host on the Mac (Homebrew mingw-w64 and the pinned toolchain), from the change's
worktree:
T=~/.rustup/toolchains/1.99.0-aarch64-apple-darwin/bin
cd host && PATH="$T:$PATH" CARGO_TARGET_DIR=<scratch>/tgt-win CODYNC_ENV=main \
CC_x86_64_pc_windows_gnu=x86_64-w64-mingw32-gcc CXX_x86_64_pc_windows_gnu=x86_64-w64-mingw32-g++ \
AR_x86_64_pc_windows_gnu=x86_64-w64-mingw32-ar CARGO_TARGET_X86_64_PC_WINDOWS_GNU_LINKER=x86_64-w64-mingw32-gcc \
"$T/cargo" build --locked --profile dev-app --target x86_64-pc-windows-gnu --bin codync-host --bin codync-hostw
The binaries land in <scratch>/tgt-win/x86_64-pc-windows-gnu/dev-app/. CODYNC_ENV=main gives
the build the installed host's identity (port 19222, ~/.codync, the CodyncHost Run value), so it
replaces that host and keeps its analytics-off setting. dev-app is a release build and sends
analytics when they're on, so only run it on the VM's host, where they're off. Without
CODYNC_ENV it would be Codync Dev (19223, ~/.codync-dev, analytics on by default).
scp both files in as codync-host-fix.exe and codync-hostw-fix.exe, then swap them in, keeping
the release's binaries:
$r = "$env:LOCALAPPDATA\Programs\codync-desktop\resources"
Get-Process Codync,codync-host,codync-hostw -ErrorAction SilentlyContinue | Stop-Process -Force
if (-not (Test-Path "$r\codync-host.orig.exe")) { Copy-Item "$r\codync-host.exe" "$r\codync-host.orig.exe"; Copy-Item "$r\codync-hostw.exe" "$r\codync-hostw.orig.exe" }
Copy-Item -Force "$env:USERPROFILE\codync-host-fix.exe" "$r\codync-host.exe"; Copy-Item -Force "$env:USERPROFILE\codync-hostw-fix.exe" "$r\codync-hostw.exe"
schtasks /create /tn CodyncHostFix /tr "`"$r\codync-hostw.exe`" serve --port 19222" /sc once /st 23:59 /it /f; schtasks /run /tn CodyncHostFix
The first start of a swapped host took several minutes to answer hello; poll it for up to ten
minutes. Until the swap is undone, the test host also starts at every sign-in. To undo it, stop
the three processes and copy the .orig.exe files back, or reinstall the release.
Reproducing Windows conditions on the Mac
When the VM can't run a check, reproduce the Windows condition in the worktree's desktop sandbox
(npm run dev -- --cdp, worktree sandboxes); the host and the
agents it starts inherit the environment:
Signed out:
CLAUDE_CONFIG_DIR=<empty folder> npm run dev -- --cdp. Use a fresh folder per run: signing in from the sheet stores a login for that folder in the keychain (remove it withCLAUDE_CONFIG_DIR=<folder> claude auth logout).A slow CLI: Claude's adapter runs
CLAUDE_CODE_EXECUTABLEinstead of its bundledclaude. Point it at a wrapper that delays one command and runs the real binary:#!/bin/sh # `claude auth status` taking 6 s, past the adapter's own 5 s limit (as on Windows). if [ "$1" = auth ] && [ "$2" = status ]; then sleep 6; fi exec ~/.npm/_npx/<key>/node_modules/@anthropic-ai/claude-agent-sdk-darwin-arm64/claude "$@"The host's own
claude auth statususes theclaudeon PATH, so only the adapter is slowed.
Limits
- Agent launches in the VM depend on its network (see agents); a check that needs an npm adapter to start may not finish.
- CI builds Windows only for pull requests and tags (
.github/workflows/host.yml); a direct push todevgets no Windows compile. Clippy the whole host for Windows on the Mac with the cross-build variables above and"$T/cargo" clippy --locked --target x86_64-pc-windows-gnu --all-targets -- -D warnings.
Investigations done this way: