Hatcheck

Several Claude accounts on one computer, each kept apart: isolated Claude Desktop instances on macOS, and Claude Code CLI profiles, with recent sessions, plan usage and resume commands across all of them. OpenAI's Codex CLI and Moonshot's Kimi Code CLI get the same treatment, Google's Antigravity CLI (agy) and Z.ai's ZCode are read beside them, and DeepSeek Harness (dsh), MiniMax Code (mcode) and Qwen Code (qwen) join as read-only agents. Sessions can be carried between the original five and Qwen Code, and out of DSH and mcode.

Unofficial: not made by, affiliated with or endorsed by Anthropic, OpenAI, Google, Moonshot AI, Z.ai, DeepSeek, MiniMax or Alibaba Cloud. Claude, Claude Code and Claude Desktop are trademarks of Anthropic; Codex and ChatGPT are trademarks of OpenAI; Antigravity and Gemini are trademarks of Google LLC; Kimi is Moonshot AI's, ZCode and GLM are Z.ai's, DeepSeek Harness is DeepSeek's, MiniMax Code is MiniMax's and Qwen Code is Alibaba Cloud's.

On macOS, hatcheck desktop runs one Claude Desktop per account, side by side: each with its own data folder and login, and a launcher in a colour you pick, which brings the running one forward rather than opening it twice.

Running Claude Desktop with its own --user-data-dir is Philipp Stracker's approach (write-up); hatcheck adds the colours and the launchers.

It also has command-line tools for all seven agents, for keeping one profile per account: cl, cx and ck run Claude Code, Codex and Kimi Code CLI with a profile, and hatcheck lists the recent sessions in all of them, with the command that resumes each one or opens DSH's session history, and what's left of each plan. The rest of the commands find, show, carry and convert sessions between the accounts and the agents, tie folders to accounts, and check how everything is installed: see the tour below, and docs/commands.md for each command as it is.

Not a gateway. Hatcheck is never between an agent and its model. It rewrites no agent's config, holds no login, and runs nothing in the background. Tools such as magpie put a local gateway in that path so any agent can use any model and accounts can take over from each other when a quota runs out. If that is what you want, use one of those. Hatcheck is for the opposite case: keeping several accounts apart, and knowing which one each session ran on. See Rules it keeps.

Claude Desktop instances (macOS)

hatcheck desktop makes them: each instance gets its own data folder, ~/.claude-instances/<name>, and a launcher in /Applications with Claude's icon in a colour you pick. The launcher brings a running instance's window forward instead of starting a second one on the same folder.

hatcheck desktop new a A --hue 190                # Claude A.app, blue
hatcheck desktop new b B --hue -10                # Claude B.app, red-orange
hatcheck desktop open a                           # or open it from Launchpad or the Dock
hatcheck desktop                                  # the instances, and which are running
hatcheck desktop remove a                         # its launcher; --data deletes its data folder too

Sign each one in to its account the first time it opens. The first time a launcher brings a window forward, macOS asks whether it may control System Events; allow it, once per launcher. --hue turns the icon's colours by that many degrees (about 110 is green, 190 blue, 240 purple) and --saturation scales them. hatcheck recolours the icon itself, with macOS's iconutil, and a launcher it can't recolour keeps Claude's own icon. The name default makes a launcher for Claude's built-in data folder, which is only safe while nothing else runs on that folder, the stock Claude icon included. docs/desktop-instances.md has the rest: what the launcher does, and how to copy a login you already have into an instance.

Claude Code, Codex, Antigravity CLI, Kimi Code CLI, ZCode, DeepSeek Harness, MiniMax Code and Qwen Code

Claude Code (the claude CLI) keeps its settings, history and session transcripts in ~/.claude (plus ~/.claude.json), and its login in the macOS Keychain, or on Linux and Windows in ~/.claude/.credentials.json. Point CLAUDE_CONFIG_DIR at another folder and you get a separate profile with its own login, so you can keep one per account: ~/.claude-a, ~/.claude-b, and so on. These are separate from the Desktop instances above, though sessions from Desktop's Code tab can be filed in them too (the session list hides those unless you pass --all). Codex works the same way with CODEX_HOME: see cx. Antigravity CLI keeps everything under ~/.gemini/antigravity-cli, which only HOME moves, so hatcheck reads that one folder: see Antigravity CLI. Kimi Code CLI works like Codex, with KIMI_CODE_HOME and one home per account (~/.kimi-code-a), which ck runs it with, and ZCode keeps its sessions in a database under ~/.zcode, which only HOME moves: see Kimi Code CLI and ZCode. DeepSeek Harness keeps sessions under ~/.dsh, moved whole by DSH_HOME, so hatcheck reads ~/.dsh, ~/.dsh-NAME and the variable's folder. DSH's own profiles are plugin stacks, not these account homes, and there is no hatcheck DSH launcher: its command is the official dsh web, whose session history you choose the session in. See DeepSeek Harness. MiniMax Code's current data home is ~/.minimax, selected by MINIMAX_DATA_DIR, or by MAVIS_DATA_DIR when that's unset; ~/.minimax-code is the installer, not a profile. Hatcheck reads its v2 SQLite catalog and canonical JSONL, and sessions resume directly with mcode --session mvs_<id>. See MiniMax Code. Qwen Code keeps everything under ~/.qwen, moved whole by QWEN_HOME, so hatcheck reads ~/.qwen, ~/.qwen-NAME and the variable's folder: each project's transcripts, which resume directly with qwen --resume <id>. See Qwen Code.

Install

One binary per platform, with no Python needed. Each of these installs hatcheck, hat, cl, cx, ck, claude-z and claude-k:

curl -LsSf https://github.com/hatcheck/hatcheck/releases/latest/download/hatcheck-installer.sh | sh
brew install hatcheck/tap/hatcheck  # Homebrew, on macOS and Linux
uv tool install hatcheck            # or: pipx install hatcheck

On Windows, in PowerShell:

powershell -ExecutionPolicy Bypass -c "irm https://github.com/hatcheck/hatcheck/releases/latest/download/hatcheck-installer.ps1 | iex"

The installer scripts put the commands in ~/.local/bin and add that folder to your PATH; set HATCHECK_NO_MODIFY_PATH=1 first to keep your shell's startup files as they are. Each release, made from a tag such as v0.2.0, has them and the binaries for each platform, and PyPI gets the same binaries as wheels. The release builds start after 0.1.0, which is the Python version.

From a clone, on Windows, switching between installs or removing one, and hatcheck doctor, which lists every copy on your PATH: docs/install.md. hatcheck installs no agent: docs/agents.md has each one's installer and app.

The commands

Each is documented as it is, with its options and what it prints, in docs/commands.md; the launchers and the agents' folders in docs/agents.md.

hat is the short name, and a subcommand can be cut to its first letters: hat m is hat move, hat r is hat resume. Antigravity CLI sessions are listed, shown and converted from its one folder, with what its /usage reports; Kimi Code CLI sessions are listed, shown, converted and moved between its homes, and ZCode's are listed, shown and converted from its database. DeepSeek Harness sessions are listed and shown from raw or compressed v4 logs and converted out, never written; dsh web opens its session history. MiniMax Code sessions are listed and shown from its v2 catalog and canonical JSONL and converted out, never written; mcode --session mvs_<id> resumes them directly. Qwen Code sessions are listed and shown from its transcripts and converted both ways, its store never written: a session converted for qwen is a file qwen is started on, and qwen --resume <id> resumes its own. claude-z and claude-k run Claude Code on Z.ai's or Kimi's models in any profile.

Rules it keeps

Each rule names the code that keeps it, so it can be checked rather than taken on trust.

• Not in the path. Nothing of Hatcheck sits between an agent and its model. cl, cx and ck set CLAUDE_CONFIG_DIR, CODEX_HOME or KIMI_CODE_HOME (and, for cl -z and cl -k, the provider variables you set up) and run the agent in their place; nothing stays running (launch.rs, run).
• No agent's config is rewritten, except when you ask: hatcheck bind --hook adds one session-start hook to each profile's settings.json, each Codex home's hooks.json and Antigravity CLI's ~/.gemini/config/hooks.json, or points the one it added at the hatcheck it finds now, and unbind --hook removes only what it added (bind.rs, add_hooks and remove_hooks).
• No credentials. It never asks for, stores, refreshes or moves a login. For Claude it never reads one: Claude Code's own /usage is run in the profile and the snapshot it saves is read (usage.rs, run_claude_usage and claude_live). For Codex, the home's auth.json token is sent only to ChatGPT's usage endpoint, as it is, and never refreshed (read_codex_login and fetch_codex_usage there). For agy it never reads one either: agy's own /usage is run, after agy models has said agy is signed in, since asked without a sign-in agy would start one (agy.rs, run_usage and signed_in). For Kimi, the home's credentials/kimi-code.json token is sent only to Kimi's usage endpoint, the one kimi's own /usage asks, as it is, and never refreshed (kimi.rs, read_login; usage.rs, fetch_kimi_usage). For ZCode, DeepSeek Harness and Qwen Code it reads none, and reads no usage: none of them has a plan-usage command or endpoint hatcheck can use. MiniMax Code reads no auth or provider config, and its plan quota isn't wired into hatcheck yet.
• Nothing leaves your computer except those usage requests: Claude Code's to Anthropic, one to ChatGPT per Codex account, one to Kimi per Kimi home, and agy's own /usage to Google, with the model list agy fetches first to say it's signed in. CODEX_USAGE_URL in usage.rs and DEFAULT_USAGE_URL and GLOBAL_USAGE_URL in kimi.rs are the only addresses the binary itself calls; DSH, mcode and qwen add none. --cached and --no-usage make none. The Z.ai and Kimi endpoints in provider.rs are handed to Claude Code, which calls them only when you run cl -z or cl -k.
• What the agents send is theirs to govern. Hatcheck adds no data path of its own. What each agent sends to its maker, and whether that is kept or used to train models, depends on the maker's terms for your product and plan, which differ between them and change; check them for each account you keep apart: Claude Code, Codex (and the API), Antigravity, Kimi, Z.ai, DeepSeek, MiniMax and Qwen Code. Hatcheck summarises none of them; the weekly check only keeps the links alive.
• No automatic switching between accounts. hatcheck best ranks and prints commands, and nothing in the binary starts, moves or switches a session on its own (best.rs: advice only, --wait sleeps and then prints).
• Nothing runs in the background. There's no daemon. hatcheck serve is the desktop app's engine and lives only while the app does, one command per line (serve.rs).
• Sessions are only ever moved, copied or rewritten in ways you can undo. move deletes the original only once every copied file is checked against it, never puts an older copy over a newer one, and is undone by moving the session back; fork keeps the original; relocate prints the command that undoes it; and all three have --dry-run, which runs every check and writes nothing (move_cmd.rs, relocate.rs). A Kimi session's line in kimi's index goes with its folder (kimi.rs, index_add and index_remove), ZCode's database is only ever read (zcode.rs, open), DSH's session logs and workspace state are only ever read (dsh.rs), mcode's catalog, canonical history and assets are only ever read (mcode.rs), and qwen's transcripts, registry and sidecars are only ever read (qwen.rs).

Function names rather than line numbers, since lines move; git grep finds them.

The desktop app

A menu-bar app (system tray on Windows and Linux) built on the CLI. It shows what's left of every plan, sorted by Fable, with bars and reset times, and the last day's sessions with Resume, or Open DSH for DeepSeek Harness's web session history, and Copy. It also has buttons for the Claude Desktop instances and notifies you when a limit runs low or resets.

Its main window browses every session, with search, filters and a read-only preview of the conversation. From there you can resume or open a session, copy its command, or move or fork it to another profile after a dry run, handing over a brief in place of the whole conversation when it goes to another agent. The window also lists profiles and homes, signs in to a new one in a terminal, and makes Desktop instances.

It ships the CLI inside itself, so it needs no Python. It never handles a login, and live usage is off until you turn it on. See desktop/README.md to build and test it, with npm run demo running it against a demo home rather than your profiles; releases come from .github/workflows/desktop.yml for macOS, Windows and Linux. The plan behind it is the proposal.

Files in this repo

• crates/hatcheck/: hatcheck in Rust, the commands and the launchers. Its src/ has a file per subcommand (sessions.rs, usage.rs, move_cmd.rs, open_cmd.rs, desktop.rs and the rest), launch.rs for cl, cx, ck, claude-z and claude-k, agy.rs, kimi.rs, zcode.rs, dsh.rs, mcode.rs and qwen.rs for what it reads and writes of Antigravity CLI, Kimi Code CLI, ZCode, DeepSeek Harness, MiniMax Code and Qwen Code, doctor.rs for hatcheck doctor, and src/bin/ for those binaries and hat. Its own README is the crate's page on crates.io. It was ported from a Python version with the same output; see docs/rust-port.md.
• CHANGELOG.md: what changed in each release of the CLI; a release's notes on GitHub are its section there.
• install.sh / install.ps1: build the commands from a clone and put them on your PATH (~/.local/bin by default).
• pyproject.toml: the PyPI package, the binaries as wheels built with maturin. dist-workspace.toml: the release builds, installers and Homebrew formula, made by cargo-dist. packaging/: the app's Homebrew cask, and the winget manifests for the app and the CLI.
• desktop/: the desktop app: Tauri, with the CLI inside it (its README).
• site/: the site, four pages sharing one stylesheet, with the images beside them, served at hatcheck.dev from Cloudflare as static assets; site/deploy.sh ships it. The logo lives there too: logo.svg and logo-dark.svg (the mark), logo-wordmark.svg and logo-wordmark-dark.svg, favicon.svg.
• docs/proposals/: the plans these came from: the CLI features, the desktop app, Antigravity CLI and Kimi Code CLI and ZCode.
• docs/: commands.md, agents.md and install.md, each command, agent and way of installing as it is; desktop-instances.md for the Desktop instances; rust-port.md on the port from Python; formats.md, each agent's session format side by side and how a conversion works; maintenance.md on checking each agent's newest version every week or two.
• AGENTS.md and docs/lessons.md: notes for coding agents (CLAUDE.md imports them): how to build and check, the rules the code keeps, and what earlier work got wrong, with the rule that would have caught it.
• tests/: the command tests, which run the binary against throwaway home folders: python3 -m unittest discover -s tests, which builds it with cargo first, and the crate's own tests with cargo test -p hatcheck. They never read your real profiles, Keychain or the network. GitHub Actions runs both on macOS, Linux and Windows (.github/workflows/tests.yml).
• scripts/: upstream.py, the read-only first step of maintenance.md, which .github/workflows/upstream.yml runs every Monday, keeping an issue open for what it finds; and provider_probe.py, which runs each claude-k and claude-z model once.

Roadmap

• Claude Desktop instances on Windows. Claude Desktop is an Electron app there too, so a separate --user-data-dir per account should work. A Windows version would create a shortcut per instance with its own icon, and a launcher that brings a running instance forward instead of opening a duplicate. It needs testing on a Windows machine first.
• Claude Desktop instances on Linux. Claude Desktop for Linux is in beta, for Ubuntu and Debian. Whether its own --user-data-dir keeps an instance apart there, as on macOS, hasn't been tried; until it has, only the CLI tools apply on Linux.
• Antigravity CLI profiles. agy keeps one login per home folder, so a second account needs a home folder of its own (HOME=~/.agy-a agy), which every command the agent runs would see too. Whether hatcheck should make such folders, and a launcher for them, is still open; until then it reads ~/.gemini/antigravity-cli only. The Antigravity app has the same one login, in the system keyring, shared with agy, so an instance of it with its own data folder would still be the same account; nothing for hatcheck to make until Google separates logins.
• ZCode. Profiles need a variable from Z.ai that moves the session database (only HOME does today), usage needs a usage command zcode doesn't have, and a session is never written into its database, which the desktop app shares: a converted session is a file zcode is started on. If Z.ai documents the import its desktop app uses or adds a zcode import command, hatcheck should use that (the proposal has the reasoning).
• DeepSeek Harness. A direct per-session open needs an official web deep link or resume command, conversion into DSH needs a supported importer, and plan usage needs a documented quota API. Until then hatcheck opens dsh web and its session history, converts only out, and never writes DSH's store. Sessions before v4 aren't read; later ones need their own reader.
• MiniMax Code. Conversion into mcode needs a supported importer rather than a model turn on an attached file, and plan quota needs a no-refresh design and maintainer approval before the regional endpoint is used. Until then hatcheck reads the v2 store, converts only out, and shows no quota data. An idle selected TUI has no durable marker, so only an active writer lease blocks opening and conversion.
• Qwen Code. A session converted for qwen is a file qwen is started on, as agy's and ZCode's are, since qwen has no importer (qwen sessions lists sessions and imports none); one written into its store would need one. Moving a session between qwen homes would copy its transcript, file history and side state, which hasn't been done. Plan usage needs a command or endpoint qwen doesn't have (its usage/ files count the tokens a session spent, not what a plan has left). QWEN_RUNTIME_DIR, which moves the session data alone and leaves the login behind, isn't followed: only QWEN_HOME, which moves the home whole.
• Releases: the CLI is on GitHub, the Homebrew tap, PyPI and crates.io from 0.2.0 (CHANGELOG.md). Next are winget, then the app's installers, once its release workflow has its signing and updater secrets (desktop.yml lists them).
• Signing: a Developer ID for the macOS app, and a certificate for the Windows one.

Credits

• Claude Desktop instances: this repo started as a fork of Aaron Oxborrow's gist, whose scripts made colour-coded launchers for isolated instances. On 16 October 2026 his scripts and his sections of this README were replaced by a rewrite, hatcheck desktop, so none of his code or text remains. The idea both implement, running Claude Desktop with --user-data-dir, is from Philipp Stracker's gist (write-up), itself a cleaned-up version of weidwonder/claude-desktop-multi-instance.
• The CLI tools: hatcheck, cl, cx, ck, the installers and the tests are by Yi Wang.
• Z.ai and Kimi: the provider settings in crates/hatcheck/src/provider.rs (endpoints, model slots, context windows and the config files) come from claude-z by Kerim (lettt325), MIT-licensed, and Yi Wang's fork of it, which added Kimi.

License

MIT, see LICENSE. It covers everything in this repo: the CLI, the desktop app, the site and the docs.