Changelog

All notable user-visible changes to metor are recorded here. Versions follow Semantic Versioning; the current version is in VERSION. Add user-visible changes under Unreleased. Ideas belong in BACKLOG.md and technical decisions in knowledge/decisions/.

[Unreleased]

• Account-enabled builds share a user sign-in and Space overview across web and native clients, with per-device hide/show settings and account-bound gateway sessions. Activation requires the coordinated account-service rollout; existing installations are unchanged by default.

• Browser restarts preserve profiles across local Space updates: Chromium stops before its display, and only dead local profile locks are cleaned up.

• API bots offer a visible model switch with remembered choices, model descriptions, startup status and specific allowance/service errors.

• Local hosted tests can create OpenCode bots using shared API test credit, with streamed chat, file reading/editing, shell commands and browser tasks. Existing subscription runtimes remain available; New Bot remembers the last access mode, runtime and model on the device.

• Server installers can use a verified preloaded image by its immutable local image ID, allowing private image delivery without registry credentials on the customer server.

• Fresh remote servers without Docker now offer first-Space setup in server management instead of a misleading timeout. Existing installations with unavailable Docker show a targeted recovery message.

[0.5.0] - 2026-09-17

• One domain for remote Spaces: new Spaces receive automatic /s/<id> addresses on the server’s existing domain. No per-Space DNS setup; pairing, browser/terminal connections and PWA notifications retain the Space path. Existing addresses remain valid.

• Multiple remote Spaces: manage isolated Spaces on one SSH server, inspect storage/RAM, create, pause/resume, update and delete with optional data retention. Remote Spaces keep running when the app quits by default; local Spaces still ask, with separate settings.

• Copilot models: update the bundled CLI to 1.0.85 (GPT-6 Astra support). Discover account-available models through the runtime, refresh the open model picker, and show when a refresh fails instead of silently presenting a stale list.

• Local Space management: choose existing Spaces or create isolated ones, inspect disk/RAM and storage paths, pause/resume, and remove containers with an explicit choice to retain or delete bot data. App connections are independent of running Spaces; quitting asks about background operation and can remember your preference.

• Browser keyboard: text follows the device’s active keyboard layout, including Mac Option, Windows/Linux AltGr, dead keys and IME. Mac editing shortcuts are translated for the bot browser; Enter inserts line breaks and held keys are released on focus loss.

• Protected Space administration: management now runs under a separate Linux account from bots, browsers and local connector programs. Existing chats, files and runtime logins are retained during startup migration; bots cannot directly read or rewrite management credentials and history.

• Bot storage: container startup moves bot configuration, chats, routines and collaboration state out of bot working folders into the management directory. Interrupted migrations resume; conflicting copies are never merged. OS-user isolation is now active. Older images require a pre-upgrade backup to downgrade.

• Connector credentials: harnesses now connect through a local host proxy for stdio, HTTP and SSE. Provider secrets stay out of generated MCP configuration and Codex arguments; old generated entries are refreshed at container boot, including paused bots. Local provider processes run with execution rights; bots sharing that account are not isolated from one another.

• Space services: Gmail tools, connection cards and browser startup/handoff now run through the bot host. Browser and metor MCP clients no longer read provider credentials or write handoff state directly. These host services now run under the separate management account.

• Bot collaboration: messages and assignments now go through the bot host. Host-bound identity, durable retries and paused-bot delivery are retained; bot tools no longer write the collaboration ledger or inboxes directly.

• Routine control: bot tools manage routines through their host, including after runtime sleep, instead of writing management files directly. Direct execution-user writes to management state are denied.

• Desktop authentication: embedded bot documents no longer inherit Space credentials. The packaged UI and explicitly opened screen/terminal frames receive scoped authorization; redirects and untrusted frames cannot reuse session cookies.

• Screen access: desktop websockify binds only to container loopback. The local wrapper publishes only the authenticated gateway; legacy desktop port mappings are removed when the container is recreated.

• Verified local installation: the macOS installer starts the Space in Terminal and waits for its gateway before reporting completion. Startup failures produce a visible error and a diagnostic report.

• Personal macOS packages: a reusable build bundles the arm64 desktop app and container image. Installation replaces the app in Applications, verifies the imported image identity, and includes startup, diagnostics and confirmed Space-reset scripts. Personal test installation removes quarantine only from the bundled app.

• Connection cards: bots can request a persistent, Space-wide Google, Microsoft or GitHub card in chat. Google supports desktop browser authorization in configured Spaces, shared Gmail search/read tools, reconnect and disconnect. Microsoft and GitHub still show setup required.

• Connection routing: Codex-native apps and remote plugin installation are disabled inside metor; unsupported dialogs interrupt the turn and cannot be followed by a misleading confirmation.

• Connection feedback: unsupported Codex authorization dialogs are cancelled with a visible explanation rather than falsely acknowledged as approved.

• Direct browser access: the bot menu opens Browser, Desktop, Terminal or Files directly. Browser shows only the page, supports keyboard/clipboard input, and cannot be covered by desktop windows. Visible login forms trigger the handoff automatically after browser navigation or interaction. Bots can also request user input explicitly; metor opens Browser, shows the request and resumes the task after Continue.

• Space notification preference: Manage Space → General can mute bot-to-bot result and blockage notifications across devices. Enabled by default; communication continues and muted alerts are not replayed.

• Bots work together across runtimes: Claude, Codex, Gemini and Copilot share the local metor tools for messaging and assignments. Results, blockages and shared-file links appear in chat; important updates use existing client/push notifications. Paused bots keep queued work, sleeping bots wake through RAM admission, and user turns take priority over queued bot traffic. Durable assignment/event records and stable inbox IDs recover interrupted deliveries without claiming exactly-once execution of external actions.

• Event routine foundation: routines can use normalized event triggers in addition to cron. Event sources write to a durable local journal; matching wakes use the existing bot inbox and are deduplicated per bot and routine. The routine tool and panel understand event triggers; provider-specific sources follow separately.

• Portable bots: export a compressed ZIP with all bot files (or none) and optional conversation context, import into local or remote Spaces, or create an independent copy. Conversations use conversation.jsonl, preserving timestamps and restoring visible chat history on import. The export dialog closes after saving. Import has an explicit file picker button and stays responsive while checking a package. Runtime credentials and sessions are excluded; imported bots and routines start paused.

• Manage server under Manage Space shows Docker diagnostics and upgrades older official releases to the app version through temporary SSH access, with confirmation and recovery on failed startup.

• Deployment tests: opt-in real Docker/Caddy recovery tests and real DNS, TCP and TLS diagnostics tests.

• More reliable server setup: interrupted installs retain extra settings and data; completed setups reconnect without restarting bots. DNS and HTTPS failures now give targeted diagnostics. Sign in with a private SSH key, including encrypted keys, as an alternative to the root password; key passphrases are never saved.

• Set up an existing server in the desktop app: verify its SSH fingerprint, sign in with a root password, review resources, then install metor and connect over HTTPS. Passwords are not saved; interrupted app installations can be retried on the same domain. First installation and use on Hetzner confirmed by the user. SSH errors distinguish rejected credentials, required password changes and connection failures.

[0.4.0] - 2026-09-11

Added

• Bots use resources on demand. Messages, routines and slash commands load only the selected runtime. Browser tools, Screen and Terminal start computer components when needed.
• Automatic sleep. Idle runtimes sleep after five minutes and wake for messages, routines or commands. Conversations and model settings persist; browser tabs stay open.
• Space RAM protection. Runtime, browser, desktop and terminal starts share an admission queue with a host-independent Space reserve. Waiting screens retry automatically; Stop cancels queued starts. Manage Space shows RAM headroom and pending components.
• Persistent local RAM allocation. Choose the new total in GiB under Manage Space. Apple containers support increases and reductions with Apply and restart, usage warnings and recovery on failure. Data and sign-ins remain on their volumes. Linux Docker supports live increases of configured limits; this path has not been verified on a Linux host.
• Event Log. Inspect, filter and export bounded diagnostic history, including sleep/wake, resource waiting, routine run IDs and processing outcomes, without chat content.
• Slash-command autocomplete. Runtime and metor entries have distinct labels and keyboard navigation. Session-aware model changes apply between messages; Codex also exposes the model’s supported reasoning efforts.
• Adaptive Screen resolution. Resizing the panel adjusts the bot’s desktop once the bot is idle. A per-device App settings switch controls this behavior.
• Screen clipboard support. Paste local text with Cmd+V/Ctrl+V or Paste, copy selected browser text, and synchronize the focused desktop Screen’s clipboard. Buttons and manual text fields remain available as fallbacks.
• Working view and document previews. Fold completed tool steps, view live thinking, and open generated documents beside the chat. Thinking is not retained in chat history.
• Multiple Spaces in native apps. Switch Spaces without reloading the interface, see unread counts across connections and open notifications in the corresponding Space.
• GitHub Copilot runtime. Sign in through the official device flow and use Copilot bots with chat, files, browser, routines and connectors.

Changed

• One Manage Space area contains the shared Space name, devices, notifications, RAM, updates and connectors. Names synchronize across devices. App settings remains separate; Remove from my overview removes only the connection on this device.
• Consistent terminology: Agent for the technology, Bot for a persistent named instance, and Space for the environment in which bots run.
• Faster local development. Build/start scripts reuse dependencies, unchanged builds and running Spaces. UI-only updates do not restart bots.
• Runtime and update information is available in Manage Space. Codex uses its runtime’s live model catalogue; existing bots keep their selected model.

Fixed

• Apple container setup installs the recommended Linux kernel without requiring interactive input, including after interrupted initialization.
• Gemini process cleanup prevents abandoned processes accumulating after model checks and bot restarts.

Upgrade notes

• Back up persistent volumes before updating. Applying a new Space image interrupts running bot tasks; bots, chats, files and sign-ins are retained. Local RAM settings are stored on the host and reused by subsequent starts and updates.
• No data migration commands are required. Open browser pages again after updating the Space; install the matching desktop build to use the local RAM editor.
• Desktop installers remain unsigned unless the release build is configured for signing. Mobile store distribution and Homebrew tap updates are separate release steps.

[0.3.0] - 2026-09-07

Added

• Pause, Resume and Run now on every routine card - they work without the bot, so a routine the quota guard paused can be switched back on when exactly that quota is used up. Creating and editing routines stays in the chat.

• Native phone app for iPhone and Android (client/mobile/, ADR-0017): the same interface in a store-ready shell, connected to a computer by its setup link, with push notifications through metor’s relay - Approve / Deny right in the notification, a switch per device, and a badge on the app icon that counts unread replies across bots and clears as you read. Bot pictures and picture attachments show in the app, and a tap on an attachment or a file in the file browser opens it with the phone’s viewer (Quick Look on iPhone). A setup or pairing link opened on a phone asks whether to sign in the app or the browser. Approve / Deny in a notification applies to the approvals that exist today, those of connectors set to ask first. Available to invited testers through TestFlight; not in the stores yet.

• A computer on a Mac can be reached from the phone in the same Wi-Fi: METOR_BIND=0.0.0.0 at metor box up publishes the interface on the network (sign-in by pairing still guards it), METOR_WATCH_BASE puts the Mac’s address into the pairing links.

• Releases carry the installer: a version tag becomes a GitHub Release with install.sh as its asset, and www.metor.com/install.sh always points at the latest release.

• An expired runtime sign-in is repaired in the chat: when Claude Code answers “Not logged in” or “OAuth session expired”, or a key is refused, the reply carries the runtime’s sign-in right there - the same steps as in the create dialog - and a bot that stopped on such an error offers it instead of Start.

Changed

• The bot list says what a bot is doing, without a status dot: three pulsing dots under the name while it starts, thinks or writes, “waiting for your approval” while an approval card is open, and when a bot stops with an error the error itself in red - it also stays in the chat as a card with the full text and a Start button. A stopped bot is greyed out in the list and gets a Start button in the header; the dot on the picture and the “ready” label in the header are gone.
• Desktop app: the Computers menu ticks the computer you are looking at (also in the tray menu) and the app remembers it - the next start opens the computer that was open last, whether it was chosen from the menu, the connect screen or by focusing its window.

Security

• A bot’s HTML or SVG file opens sandboxed: its scripts no longer run with your session and cannot call the interface’s API. Files the bots write are also served with their real path only
  • a link a bot leaves in its directory cannot lead outside it - and the connector file with its keys is no longer listed in the file browser.
• Sessions end after a year on the server too, not only in the browser’s cookie, and removing a device now closes its open screens, terminals and live updates at once. Changes with the session cookie must come from the interface’s own address (cross-site requests are refused).
• Desktop app: microphone, camera, screen sharing and the clipboard are granted to the interface itself only; screen sharing needs a click. Desktop and phone app refuse a computer address with plain http:// outside this machine and the local network - the session would travel unencrypted.
• The server compose file caps the box at 8 GB (METOR_MEMORY) and 4096 processes.

Fixed

• Messages sent while the bot was busy survive a restart of its host: they were lost when the host process died before it got to them, although they stood in the chat.
• A routine with a schedule that never matches a date (the 31st of February) is refused instead of being given a made-up time; one that runs out of dates pauses with a note.
• Two uploads in the same second no longer overwrite each other.
• The installer’s compose file now comes out of the image, so it always carries the volumes the image needs (the Gemini login volume was missing from the installer’s copy).

[0.2.0] - 2026-09-05

Added

• A picture per bot: every bot shows a picture in the list and the header. By default it is one to three initials on a colour - the create dialog fills the initials in from the name as you type and picks a colour, both can be changed there - or an image of your own (PNG, JPEG, WebP or GIF up to 2 MB) chosen right in the dialog. A click on the picture in the header changes all of it later.

• Gemini CLI as a third runtime (ADR-0016): bots on Google’s Gemini, signed in from the create dialog with a Gemini API key from Google AI Studio - the free tier needs no subscription; the key is checked once and stays inside the bots’ computer. Gemini bots chat, use their computer and browser, run routines and send files like the others; connectors from Settings apply to them too.

• The newest models, always: Claude Code bots can pick Fable next to Opus, Sonnet and Haiku - each name means the newest model of that family, so the list keeps up with Claude Code itself. The names carry the version (“Fable 5.1”, “Opus 5”): first from what Claude Code reports, then from what the bots’ own answers actually ran, which is what counts. Codex bots see the models Codex offers right now (asked from Codex, not a fixed list). The create dialog’s Other model id… (and metor bot create --model) takes a full model id such as claude-fable-5-1.

• metor on your own Mac: metor setup picks a container runtime (Apple’s container on an Apple silicon Mac with macOS 26, otherwise Docker or Colima), gets the image, starts the bots’ computer and opens the setup link - no server, no Docker Desktop. The desktop app does the same from its connect screen (“Set up the bots’ computer on this Mac”, with the progress shown live) and starts a stopped local bots’ computer whenever the app opens. The published image now comes for amd64 and arm64.

• Connect in two steps: the app first asks where the bots’ computer should be - on this Mac or on a server, one sentence each - and only then what that needs. On this Mac it acts on its own: none yet → set up, stopped → start, running → open, several → pick. A bots’ computer that is already open in a window never gets a second one: the app jumps to that window.

• Whose computer: the interface now always says “the bots’ computer” (and “this Mac” for your own device) where the two could be confused - the sign-in page, the connect screen, the app’s menus and the error messages.

• Desktop app for macOS, Windows and Linux (client/desktop, ADR-0015): the same interface as an app of its own - connect with a setup link, a pairing link or a pairing code, several bots’ computers side by side, native notifications for approvals, finished replies and unexpected stops while the app runs, a tray icon, the bot’s screen and terminal as before. The app shows up under Settings → Devices as “metor app on Mac/Windows/Linux” and can be signed out there.

• Sign-in for apps and scripts: the session secret can travel as a bearer token (Authorization: Bearer …) instead of the cookie; a setup or pairing claim can be redeemed as JSON (POST /bots/api/auth/redeem); GET /bots/api/version tells version and capabilities before signing in.

• Routines next to the chat, in plain words: a calendar button in the header shows the bot’s routines as cards - “Weekdays at 07:00”, next and last run, active or paused with the reason, the task - plus the recent runs; no cron expressions or ids in sight.

• One computer button instead of three views: the bot’s screen is shown next to the chat with a button in the header, and the divider between chat and computer can be dragged (the share is remembered per device). On a phone the computer replaces the chat while it is shown.

• Ticks instead of “delivered”: sent messages show one tick while on the way, two grey ticks once the bot has taken the message and two green ticks once it has answered.

• Messenger-style bot list: every bot shows an avatar with its status dot, the time of the last message (the time today, “Yesterday”, the weekday within a week, otherwise the date), a one-line preview of that message (“You: …” for your own) and a badge with the number of unread bot messages. Opening a chat marks it read on every device. “Compact bot list” in Settings → Appearance switches to one line per bot.

• Connectors (ADR-0014): under Settings → Connectors you add MCP servers that every bot can use - from a small directory (Memory, Sequential thinking, Context7, DeepWiki, GitHub, Brave Search, Exa, Firecrawl, Hugging Face, Stripe) or as a custom connector with a command or a URL, environment variables or headers. Both runtimes get them; a connector is live once the bot has (re)started, and one button restarts the running bots.

• Messenger header: the bot list starts with the metor wordmark, the ⋮ menu (Settings) and a round + for a new bot, like a messenger; the buttons at the bottom are gone.

• Settings replaces the Devices button (now in the ⋮ menu): one dialog with the sections Devices (sign-in, pairing, notifications), Appearance (text size, roles in the bot list) and Behaviour (bot list sorted by latest activity - newest chat on top like a messenger; the view a bot opens with). Appearance also decides when the Claude quota bar shows: always, never, or only from a chosen usage (for example 80 %). Appearance and Behaviour are remembered per device.

• Free-text bot names: a bot can be called Fußball-Späher 2 or Mein Bot!. The create dialog derives the id (fussball-spaeher-2 - directory, links, address between bots) live and lets you change it; the name is what the sidebar, the header, notifications and the bot’s own instructions show. CLI: metor bot create "Mein Bot!" [--id mein-bot].

• Phone app with push notifications (ADR-0013): the interface installs as an app (Android: “Install metor as an app” under Settings → Devices or the browser menu; iPhone and iPad: Share → Add to Home Screen, then sign in once inside the app with a pairing code). Turned on under Settings → Devices → Notifications on this device, it sends a push when a bot needs an approval, has finished a reply or stopped unexpectedly – never to the device that is looking at that chat. No third-party service in between: the box signs and sends the messages itself (needs HTTPS).

• Claude Code sign-in from the interface: the “New bot” dialog now runs the official Claude login too (open the link, sign in, paste the code the page shows) - no terminal needed. Codex keeps its device-code flow.

• Sign-in without passwords (ADR-0012): the first browser gets in with a one-time setup link (printed by the installer and by metor auth link), further phones and computers are linked from a signed-in device by QR code, link or pairing code (valid two minutes), and the new Devices dialog lists every signed-in browser and signs one out. The gateway checks the session on every request, including screen and terminal.

• Unattended installation: install.sh installs Docker by itself when it is missing (METOR_INSTALL_DOCKER=no stops instead), reads the domain from METOR_DOMAIN and the ghcr login from METOR_GHCR_USER/METOR_GHCR_TOKEN when set, and asks only for what is missing - one ssh line installs a server; it ends with the setup link.

Changed

• The bundled Caddy configuration no longer carries a login (no user, no password hash, no WebSocket exception) - Caddy only terminates TLS; the installer asks for the domain only.
• Routine limits: up to 30 routines per bot, names up to 60 characters, the last 30 runs are kept. The routine that pauses itself after many unattended runs is now called “auto-pause”.

Removed

• The old background-session mode (harness: "claude", Claude Code --bg with Remote Control and the Haiku chat bridge) is gone; every bot now runs through its runtime’s host process. Bots that were still on the old mode keep their role and files but start a fresh session.

Fixed

• After a restart of the computer a bot could show as “stopped” for a long time although it answered (its host mistook a stale process ID for a running twin).

[0.1.0] - 2026-09-02

First public release. Everything below was built between 2026-08-30 and 2026-09-02.

Added

Bots and their computer

• Computer (the box) with one desktop per bot: virtual display, Chromium, terminal window, noVNC access; the desktop has a dock to switch between browser and terminal.
• Bots as persistent sessions with name, role, memory and their own directory. Bots can assign tasks to each other (Claude bots) and send watch links (with token) when they need a hand in the browser.
• Bots run as Agent SDK sessions (ADR-0009): the chat shows answers, tool activity and clickable approval cards; a restart resumes the session and its context.
• Codex as second runtime (ADR-0011): bots run on Claude Code or Codex (ChatGPT subscription). Runtime and model are chosen when the bot is created (also for Claude bots: Opus / Sonnet / Haiku). Codex bots chat, use shell and files, drive their own browser, create routines and send files to the chat; restarts keep the context.
• Setup wizard: a runtime that is not signed in yet shows the official device-code login in the “new bot” dialog (link + one-time code, confirmation on the phone). Login data stays inside the box (new volume metor-codex). A login failure of one runtime stops only that runtime’s bots.

Interface

• Own web interface under /bots/ (ADR-0008): create and control bots, chat, watch the desktop (noVNC), Markdown in the chat.
• Token streaming: answers type in live (a “writing” bubble with Markdown).
• Stop button for running turns, expandable tool results, quota display (5 hours / week) for the Claude subscription in the sidebar, optional ntfy push when a bot waits for an approval.
• Terminal tab: a real terminal (ttyd / xterm.js with copy and paste) as a sub-tab of the computer panel, per bot, behind the gateway.
• File browser: third sub-tab “files” in the computer panel - navigate the bot’s directory, view and download files (breadcrumb, dot directories hidden).
• Attachments in the chat: paperclip, paste screenshots (Cmd/Ctrl+V), drag and drop; chips with a remove button above the input, thumbnails in the history. Files are stored under <bot>/uploads/; the bot reads images itself (runtime-neutral). Max. 25 MB per file, 10 per message.
• Files from the bot: when the bot writes [File: path] in its reply, the file appears as a card (image preview or download) inside its message; file paths mentioned in the text are offered as cards automatically as well.
• Responsive interface: messenger pattern on phones (bot list -> chat with back arrow and back gesture), actions in a “more” menu, “side by side” only at desktop width.

Routines

• Routines (ADR-0010): bots create their own schedules via chat (add_task / list_tasks / remove_task); the supervisor fires them in box local time and catches up a missed run once; routines panel in the interface. Runtime-neutral.
• Routine polish: update_task (change / pause / resume) and an auto-pause - a routine pauses after 20 runs without user activity (configurable via METOR_ROUTINE_GUARD).

Installation and operations

• Server installation: prebuilt image on the GitHub Container Registry (built by a GitHub Action), deploy/compose.yml with or without bundled Caddy (TLS and login), one-line installer curl ... install.sh | sudo bash, INSTALL.md with all .env options and deploy/.env.example. Reference deployment behind a Caddy login.
• This changelog.

Changed

• Mono-repo with backend/, frontend/ and knowledge/; product name metor (ADR-0007).
• Agent SDK sessions are the default; the previous background-session mode remains as a fallback.
• Bus log rotation.

Fixed

• Screen and terminal behind the server login: WebSockets failed on Basic Auth (browsers send no credentials there) - the gateway now checks a watch cookie instead.
• The screen stayed on “connecting” because x11vnc hung on an absurdly high Docker nofile limit; the container now runs with 65536.

This page is metor-com/metor/blob/main/CHANGELOG.md in the metor repository, rendered as it is.