Skip to content

v0.2.0

Choose a tag to compare

@github-actions github-actions released this 01 Nov 09:26
· 124 commits to main since this release
76f085b

Added

  • ACP agents (ADR-47): the daemon drives agents that speak the Agent Client Protocol (v1), such as OpenCode, Gemini CLI, Goose, Kimi Code, Qwen Code, Cursor, GitHub Copilot, and Pi, through one adapter. Each one is its own entry in the agent list and is configured under [agents.acp.<name>]; the agents above have built-in commands, and any other ACP command can be added. An agent counts as installed when its command is on the daemon's PATH; npx / uvx launches are never run. One process serves all of an agent's sessions, and sessions resume with session/resume or session/load when the agent supports them. Models, thinking levels, modes, and other options come from the agent's configOptions; mark risky modes with dangerous_modes. The daemon does not offer file system or terminal access to ACP agents: they use their own tools and ask through the usual permission cards.
  • Web UI: full-text search of session content. The ⌘K palette searches user messages and assistant replies on every selected online node through the daemon's index, next to the local title matches; picking a hit opens the session at that message and highlights it, loading older history if needed. Snippets are shown as plain text with only the matched words highlighted, and the daemon no longer keeps the control characters its snippet markers use in indexed text.
  • Web UI: read-only file browser at /n/:node/files/…, opened from a project page (Files) or from a session's working directory in the info panel. It has a lazily loaded tree, highlighted source, rendered Markdown (relative links and images open files on the node and are never requested from the hub), and image previews; text files show their first 1 MiB. Protocol: fs.read takes encoding: "base64" and offset, and returns raw bytes in chunks of at most 512 KiB (data_base64, with the offset echoed), so a large image does not hold up other messages to the device.
  • Remote node settings (ADR-43): node.get_config / node.set_config read and change a whitelist of hot-reloadable settings: concurrent sessions, idle and permission timeouts, notifications, log level, worktree defaults, and the preview ticket lifetime. permissions.allow_dangerous_modes and preview.enabled can only be turned off remotely. Agent commands, directories, session env, preview ports, and the hub address can only be changed on the machine, and a setting fixed by an ANYCON_* environment variable is read-only. A request is checked as a whole, written with the file's comments kept, and audited per key. node.logs returns the tail of the daemon log, leaving out lines that may quote agent output. Web UI: Node settings → Limits is an editable form, and About has View logs.
  • Device token rotation: control.rotate_token / token_rotated. The hub issues a new device token and accepts the previous one for 60 seconds. A device rotates at most once a minute, and only from a connection that authenticated with its current token, so a leaked older token cannot be used to take the device over. The daemon, the Web UI, CLI devices, and the TS SDK (for identities it stores) rotate a token once it is 30 days old. SDK: client.rotateToken(), the tokenRotated event, and onTokenRotated. Database: migration 0004_token_rotation.sql.
  • Hub: Postgres storage (ANYCON_DATABASE_URL=postgres://…, ADR-12). It behaves like SQLite, with row and advisory locks where SQLite relied on having a single writer; both backends share one test suite, and their migrations share numbers. docker/compose.postgres.yml is an example setup. Keep the /data volume with Postgres too (Web Push keys, admin token) and back up the database with pg_dump. The hub still runs as a single instance. See docs/deploy.md § 3.2.
  • Multiple workspace roots (ADR-45): workspace.extra_roots adds folders whose subfolders are projects, next to workspace.root (the primary root, which keeps .scratch). Project names stay unique across roots: linked projects win, then the primary root, then the extra roots in order. Overlapping roots, or a root inside the data directory, stop the daemon from starting; a root that is missing (an unplugged disk) is skipped with a warning and reported by anycon doctor. Each root has its own trash. Changing the roots needs a restart. Protocol: node.info.workspace_roots, Project.root, TrashEntry.root, and an optional root on workspace.create / workspace.clone. Web UI: a Location choice when creating or cloning a project. CLI: anycon projects create|clone --root.
  • anycon export / anycon import (ADR-46): back up session history and attachments to a zip file, or move them to another machine. Only the local daemon does this; remote clients cannot. A plain anycon export includes everything, archived sessions too; --project, --since, --session, and --exclude-archived narrow it. The file is written with owner-only permissions and contains conversation content in plain text. Import checks the archive strictly, skips sessions that already exist, imports each session in one transaction, and rebuilds the search index. It does not keep the agent's own session id, so the next message in an imported session starts a new agent conversation while the history stays. --dry-run checks without writing. The file format is specified in docs/spec/daemon.md § 20.
  • Attach any file, not just images (ADR-44). Images still go to the model as images when the agent takes them natively. Every other attachment is copied by the daemon into the session's working directory (<cwd>/.anycon/attachments/<session_id>/, owner-only, created through directory file descriptors on Linux and macOS so a symbolic link swapped in by the agent cannot redirect the write), and the agent gets the relative paths, types and sizes at the end of the message text. So every agent can use every file type, and none has to ask for permission to read outside the workspace. /.anycon/ is added to the repository's info/exclude, so the files never show up in git status. The conversation history keeps what you wrote, without that list. If the copy fails (disk full, a symbolic link in the way), the turn ends with an attachment_failed error and nothing reaches the agent; fix the cause and send again. Deleting a session removes its directory. Protocol: blob.put_begin accepts any mime (lowercased, parameters dropped; an empty or malformed one is recorded as application/octet-stream); only the four image types get a header check and width / height. capabilities.attachments lists the mimes an agent takes natively. node.info reports capabilities.file_attachments and attachment_limits. The per-file, per-message, and quota limits are unchanged. The Web UI adds files through the "+" menu, paste, or drag and drop (a file over the node's per-file limit is refused before it is read; against an older node it still takes images only). Files show as chips that download on click and are never rendered inline. CLI: anycon send --file <path> and anycon new --file <path>. SDK: node.uploadFile() and acceptsAttachment(node, agent, mime); uploadImage() refuses a non-image mime locally. A title or preview built from attachments says [file] when any of them is not an image. On Windows the write is still checked first and then done by path, so a link swapped in at that exact moment is not caught.
  • Windows support for anycon (daemon and CLI, x64, Windows 10 1809 / Server 2019 or newer; ADR-42). The local CLI channel is a named pipe owned by and restricted to the current user, and the CLI checks the pipe's owner before talking to it. Agents and the processes they start run in a Job Object, so stopping a session or the daemon leaves nothing behind. ~/.anycon gets an owner-only ACL. anycon service install registers a per-user scheduled task that starts the daemon at logon without a console window (through the new anyconw.exe launcher) and needs no administrator rights; service stop asks the daemon to shut down cleanly first. Release archives include anycon-x86_64-pc-windows-msvc.zip, installed with irm …/install.ps1 | iex. Environment variables are matched case-insensitively on Windows, so a session env key like path can no longer stand in for PATH. Windows-specific system and profile directories (AppData, PowerShell profiles, Program Files, …) are on the linking deny list. Preview port detection uses the system's TCP table and the agent's job instead of lsof; a dev server started outside a session needs preview.allow_ports there.
  • Web UI: the setup wizard has a Windows tab with the PowerShell installer.
  • Background tasks, notices, and transient activity in the event stream (ADR-41). Protocol: new task, notice, and activity events. task is a snapshot of a workflow, background subagent, or background command (a later snapshot with the same id replaces the earlier one); a workflow lists its phases and each agent's state, tokens, last tool, and error. The daemon sends at most one snapshot per task per second (status changes are never delayed) and marks tasks that were still running when the agent process went away as cancelled. notice carries non-error messages from the agent (model fallback, context compacted, rate limit, local command output such as /context, config and hook warnings, …) and never raises a push notification. activity shows what the agent is doing between outputs (retrying an API request, compacting context) until content arrives. Clients must tolerate notice.code values they do not know.
  • Protocol: turn_done.usage.context_window and context_tokens (context window of the model, and tokens in it at the end of the turn), and detail.url on kind = other permission requests (a link the user is asked to open in a browser). All fields are optional.
  • CLI: anycon attach / new / send print task starts and finishes, notices (a local command's output as is), and the retry / compaction status, and show the full URL of a permission request that asks to open a link. --auto-approve never answers such a request.
  • Mock agent: the prompt keyword tasks emits activity, a notice, and a workflow task linked to a Workflow tool call; link asks to open a URL. turn_done.usage carries the context window and tokens used.
  • Protocol: Question.optional marks a question in a kind = question request that may be left unanswered; the default (required) keeps AskUserQuestion as it was. A completed task can go back to running (a Codex sub-agent that receives new input), so clients treat the latest snapshot as the truth.
  • Claude Code adapter: a workflow (including the Workflow tool) and background subagents and commands report their progress as tasks, with phases and each agent's state, tokens, and last tool. API retries and context compaction show as a status while the turn runs, and the context window and tokens used come with each finished turn. Notices cover a model fallback, a usage-limit warning, hook and MCP server failures, denied tool calls, and the output of local commands such as /context and /cost. An MCP server that asks for input (elicitation) is no longer cancelled automatically: its form becomes a question (fields the server does not require can be left blank) and its sign-in link becomes a permission request with the full URL.
  • Codex adapter: a sub-agent is shown under the call that started it, with its state, tokens, last tool, and result, and its text and tool calls stay out of the main conversation. A running sub-agent counts as a background task, so the agent process is not reclaimed while it works. Notices cover configuration and deprecation warnings, MCP servers that failed to start, a rerouted model, declined automatic approvals, and hook failures. Context compaction shows as a status and a notice. MCP elicitation forms and links are shown like the Claude Code ones, and approving an MCP tool call offers Codex's own choices (this time, this session, always, or deny).
  • Web UI: the progress of a background task sits right under the tool call that started it: a workflow lists its phases and agents, a subagent or command shows its state, tokens, and last tool. The tool line keeps showing "running in background" until the task ends, even when the call itself returned at once. A chip above the input says how many background tasks are running and lists them. Notices appear as lines in the conversation (local command output as its own block), a retrying or compacting status line shows while the turn runs, the info panel shows context usage, and permission cards show the full link with an Open button. Optional questions say so and can be skipped. Exported Markdown includes notices and the final state of each task.
  • Answered questions stay in the conversation: each question is listed with its answer, and a note sent with a reply (a skip note or a denial reason) is shown under it. Protocol: permission_resolved carries the device's answers and message. Both fields are optional, and replies recorded before this change show the one-line summary as before.
  • Claude Code background commands (run_in_background) show "running in background" until they finish, then get their final result and exit code. The session panel's Agents card lists them while they run and counts them as working, along with any other background tasks the daemon reports.
  • Web UI: a "Recent" group at the top of the sidebar, under "Needs approval", lists the sessions that are running (running, starting, or queued), newest first, each with its project, so you no longer have to look through every project. Sessions that need approval stay in their own group and are not listed twice. Turn it off under Settings → Appearance → "Recent sessions in the sidebar" (on by default; the choice is kept per browser).
  • Web UI: add a project without going to the node's home page. Use the new button next to search in the sidebar, "New project…" at the end of the project list on the new-session page, or "New project" / "Clone a repository" / "Open an existing folder" in the ⌘K palette. The dialog can switch nodes, and a project created from the new-session page is selected there.
  • Web UI: Node settings → About has "Remove this node", for a machine that was reinstalled or retired. It works while the node is offline: the hub revokes the node and this browser forgets it. Revoking a node under Settings → Devices now forgets it too. Other browsers list a node removed this way under "Removed from your account" and can forget it. Projects and session history on the machine itself are not touched.
  • Dev server preview (ADR-39): open a dev server running on the daemon machine from the session page, embedded or in a new tab. The page loads on a separate preview domain (ANYCON_PREVIEW_URL, https://*.aiplace.app on the public hub) and its traffic goes through its own Noise tunnel to the daemon, so the hub still only forwards ciphertext. By default the daemon only forwards to ports opened by an agent session; other ports go in preview.allow_ports. Protocol: preview.* methods, auth.preview_token, preview_peer, and an optional to on rate_limited. Node settings gain a Preview tab that lists and closes previews. Database: migration 0002_preview.sql.
  • Hub multi mode (ANYCON_MODE=multi, ADR-40): open registration, the configuration for a public hub. The first email-code or GitHub / Google sign-in creates the account. There is no bootstrap token, token login, or invitation, and the hub refuses to start unless SMTP or an OAuth provider is configured. Admins are the accounts whose email is in ANYCON_ADMIN_EMAILS, plus accounts an admin has promoted. An existing single database can be switched to multi. See docs/deploy.md § 9.
  • Per-account quotas in multi mode: devices, daemons, simultaneous connections, and pairing / enrollment attempts per rolling 24 hours (ANYCON_QUOTA_DEVICES, _DAEMONS, _CONNECTIONS, _DAILY_PAIRINGS; defaults 10 / 3 / 10 / 20). Every completed match costs exactly one slot: starting a pairing or enrollment counts once (pair_begin, enroll_begin), and when the node that started the pairing has no account yet (a first-time daemon), the account that pairs it pays the slot instead. A mistyped code costs nothing, and an account cannot pair fresh daemons all day. A pairing that would go over fails with pair_failed { quota } / enroll_failed { quota } and creates no device, and a connection over the limit gets auth_fail { quota }. Protocol: auth_ok.quota and quota_update carry the usage. The hub also serves GET /api/quota, and the Web UI has Settings → Quota.
  • GitHub and Google sign-in (ANYCON_OAUTH_GITHUB_CLIENT_ID / _CLIENT_SECRET, ANYCON_OAUTH_GOOGLE_…), in both modes. Authorization code flow with PKCE and a state cookie; the callback only works on ANYCON_PUBLIC_URL, so the reverse proxy must pass the original Host (a request on another host is redirected to the public address at most once, marked hop=1, so a proxy that rewrites Host cannot cause a redirect loop). Failures, including a hub-side error (server_error), redirect back to the Web UI with a stable oauth_error code. In single mode new users sign up through OAuth with an invitation. Link and unlink accounts under Settings → Login methods (GET /api/me/identities, DELETE /api/me/identities/:provider). An email the provider has not verified is never used to match or create an account. The hub stores only the provider and its stable user ID, not names, emails, or tokens.
  • Email-code sign-in in multi mode: the hub sends a code to any valid address, and entering it creates the account if there is none yet.
  • Delete your account: Settings → Profile (DELETE /api/me, needs a sign-in from the last 10 minutes). The hub removes the account's devices, pairings, sign-in methods, and audit records, disconnects it immediately, and keeps a single account_deleted record. Session history on your nodes is not touched.
  • Admin page: search accounts (substring of the name or email, or an exact ID; ASCII letters are case-insensitive, non-ASCII characters in a display name must match in case), grant or remove admin, and delete accounts (GET /api/admin/accounts?q=, POST / DELETE /api/admin/accounts/:id/admin, DELETE /api/admin/accounts/:id). Invitations are hidden in multi mode.
  • Links to your terms of service and privacy policy (ANYCON_TERMS_URL, ANYCON_PRIVACY_URL, both modes, https only). When set, /api/config carries terms_url / privacy_url, the sign-in page shows "By continuing, you agree to the Terms of Service and acknowledge the Privacy Policy" with links that open in a new tab (only the document that is configured is mentioned), and anycon connect prints the same line before it sends an email sign-in code. Nothing changes when neither is set. A hub with open sign-up (multi) should set them; the public hub uses https://anycon.ai/terms and https://anycon.ai/privacy.
  • Database: migration 0003_identities.sql adds the identities and oauth_states tables. Migrations only move forward, so back up /data before upgrading if you may want to roll back.

Changed

  • Web UI: unpinning the wide-screen sidebar (the pin button next to the logo, or ⌘/Ctrl+B) leaves a narrow strip with search, new project, the number of sessions waiting for approval, and a pin button. Hovering the strip slides the sidebar out over the page, and it slides back once the pointer leaves, on a click outside, or on Esc. On touch screens, tap the strip. The "show sidebar" button in the page header is gone.
  • login_methods in /api/config and /api/me is ordered github, google, email, token, so a single hub with SMTP now reports ["email","token"]. The CLI and Web UI do not depend on the order.
  • ANYCON_ADMIN_EMAILS, ANYCON_QUOTA_DEVICES, _DAEMONS, _CONNECTIONS, _DAILY_PAIRINGS, and ANYCON_OAUTH_* no longer log "recognized but not implemented". Blank values count as unset. A value that is set but invalid stops the hub from starting, in both modes. In the mode that does not use a variable it is ignored with a warning, and ANYCON_ADMIN_TOKEN is ignored in multi.
  • Removing an account's last way to sign in is refused with 409 last_login_method: unbinding the email, unlinking a GitHub / Google account, and (in single) deleting a login token. The three are serialized per account and judged on the account as it is inside the lock, so two requests at the same time cannot both succeed and leave the account locked out. The single bootstrap admin always keeps the bootstrap token.
  • Web UI: the message list and the input share one column. Its text width follows the center pane, up to 72rem, with side margins that grow from 16px to 80px as the pane widens. User message bubbles are allowed up to 56rem.
  • Web UI: on wide screens, the "will be sent after this turn" notice sits on the same row as the branch chip, on the left.

Fixed

  • Web UI: the sidebar no longer lists a project the node no longer has (moved to trash, unlinked, or its folder deleted). Its sessions are still on the node and open from the node's home page, search, and the project page.
  • Web UI: the "Incompatible versions" and "This node no longer recognizes this device" pages have a "Node settings" button and no longer block the node's settings page, so a node that may never connect again can still be removed from the account.
  • anycon pair on an already paired node stops as soon as the hub refuses the pairing (for example quota when the account's device limit is full, or the new device disconnecting) and says why, instead of waiting until the code expires and reporting a timeout. It also stops, saying so, when another pairing was started on the same node (the Web UI's Add device, a second anycon pair) and replaced its code. A refusal when the code is registered (the daily pairing limit, rate limiting) is no longer reported as the node being offline: node.pair_begin returns busy with data.reason quota / rate_limited (or forbidden), unsupported stays "not connected to the hub", and the Web UI's Add device dialog explains the quota case with a link to Settings → Quota. --json prints {"failed":{"reason":…,"message":…}} as the last line. The node keeps the outcome of the pairing it started and exposes it through the local-only local.pair_outcome.
  • Web UI: the session info drawer no longer pops up by itself when the window is widened past 1200px and narrowed again.
  • The daemon reconnects when its hub connection goes silent for 75 seconds. Before, a proxy or TUN that dropped the connection without closing it (for example Clash reloading its config) left the daemon thinking it was online while clients showed the node as offline, until the daemon was restarted.

Install

Hub image (linux/amd64, linux/arm64):

docker pull ghcr.io/anycon-ai/hub:0.2.0

anycon CLI and daemon (macOS arm64 / x64, Linux x64 / arm64 with glibc 2.35+):

brew install anycon-ai/tap/anycon
# or
curl -fsSL https://github.com/anycon-ai/anycon/releases/download/v0.2.0/install.sh | sh

Windows x64 (PowerShell):

irm https://github.com/anycon-ai/anycon/releases/download/v0.2.0/install.ps1 | iex

npm: @anycon/protocol@0.2.0, @anycon/sdk@0.2.0.

Each archive contains THIRD-PARTY-LICENSES.txt. Verify an archive with sha256sum -c SHA256SUMS or gh attestation verify <file> --repo anycon-ai/anycon.