mistgate Docs
EN
316

Status and roadmap

What Mistgate releases already do, what is being tested now, what comes next and what 1.0 needs, and the known limits of the current release.

On this page

Mistgate is in beta; the current release is v0.1.85. It runs in production, but the API, the stored settings and the node protocol may still change before 1.0. The rest of these docs describe only what the code does today; the plans are on this page. The plans have no dates: "Next" and "Later" give the order, not a schedule.

  1. Released#

    v0.1.0 – v0.1.85

    Everything below is in a release and in the docs. First what the current release brought, then the rest by admin section.

    • New in the current release#

      v0.1.85
      • The Updates page in three layers, and the servers' operating system. Three cards at the top: This panel, Node agents and Server systems. The new layer only reads: the distribution, the kernel, waiting and security updates, whether a reboot is needed and whether the AmneziaWG module is built for the next kernel; installing updates and rebooting from the panel are not there yet. Needs the agent v0.1.85. See Server operating system.
      • A resumed update download takes the slot at once. When a node's link to the panel stalled, the resume waited and was refused, and on a bad link the minutes of the window were lost. A resume of the same file now takes the node's slot immediately. See Where the bundle comes from.
      • GitHub as a fallback update source. A node can finish an update from several sources in turn; the panel offers GitHub second when the release there is signed by its key and carries the same files. On by default, switched off in Settings → System → Download sources; GitHub sees the nodes' addresses. Needs the agent v0.1.85. See Download sources.
      • An hourly limit on user changes for API tokens. A token without the owner may take access away from, or rewrite, at most 10 people an hour, so a leaked token cannot switch everyone off at once; creating, extending and enabling people do not count. The owner lifts the limit with a switch on the token; existing tokens are limited after the update. See Limits.
      • A node's own ports from the tunnel. A full-tunnel AmneziaWG client now reaches the node's other published inbounds (a Telegram proxy, VLESS and Hysteria2 on other ports), not only TCP 443. Needs the agent v0.1.85. See Troubleshooting.
      • Groundwork for Mistgate's own app. The panel has a client API: the owner's switch, enrolling a device by a key, sync and leaving. It is off by default, and nothing changes for existing apps.
      • Security: no Multipath TCP on our listeners (v0.1.83). Go opens every listener with MPTCP by default, and CVE-2026-98432, a flaw in the Linux kernel's TCP (CVSS 9.8), was shown as remote root through an MPTCP listener. The panel and the node agent, with the Xray listeners it hosts, now open plain TCP; update the panel and the nodes. The bug itself is in the kernel: also set net.mptcp.enabled=0 and install your distribution's patched kernel when it is out. See Troubleshooting.
    • Profiles#

      • Hysteria2 on the official core with Salamander.
      • AmneziaWG 2.0 and 3.1 in userspace or as a kernel module, several profiles per node.
      • VLESS on xray-core: REALITY over XHTTP or TCP (RAW + Vision), with the node's own site or another site as the target; TLS over XHTTP or WebSocket, direct or behind a CDN.
      • VLESS and Hysteria2 on one node, TCP 443 included.
      • MTProto, experimental: a proxy for the Telegram app with Fake TLS, one key per person.
    • Nodes#

      • SSH installation from the admin, and through MCP with the owner's approval: a pinned host key, checks before any change, jobs that can be cancelled, retried and resumed after an interruption.
      • Enrolling a node by command: mistgate-node enroll explains a refused TLS handshake (a CDN or a reverse proxy in front of the panel) and names the flags that fix it.
      • Saved SSH access, encrypted, with a verified password change; the panel's own SSH key, the server's authorized_keys from the node card, and access for nodes installed by command.
      • WARP egress that fails closed when the WARP link drops, and torrent protection on every node.
      • Check ports: UDP delivery and a TCP connection to every port of the node from another node of the fleet.
      • Measuring the network capacity of a node's link.
      • A decoy site per node: the built-in one, the panel's site or an archive of its own.
      • Relays through an entry node: checked before they are switched on, watched every 5 minutes, a copy of the server in subscriptions; AmneziaWG servers too, with a key of their own for the path through the entry.
      • IPv6 entry: the node's servers at one of its IPv6 addresses, checked from outside before they are turned on.
      • Resources and load: what a node and Mistgate itself cost the server, a Load view with the room left for more people, a debug capture of one node, a map of the fleet and the ways in as a graph.
    • Users#

      • Users and groups: traffic quotas with periods, terms, devices by device ID with a limit in three modes, speed per person and per device; a group also decides which nodes its people reach.
      • The People tab sorted by any column, with device counts, and bulk actions on several users at once.
    • Subscriptions#

    • Health#

      • Client-eye checks: every 5 minutes the panel connects to the servers the way an app does, VLESS included (REALITY and TLS, directly or through the CDN).
      • Connection counters from the nodes: connections, drops right after connecting, refused keys, handshakes that never complete, time online and unknown keys, per device on a person's page when attribution is available and per profile on the node. Hysteria2 and AmneziaWG event counters, plus VLESS session time and, with an agent from v0.1.79, unknown keys; for Hysteria2 also how connections end (a frozen path, a quiet client, a close), RTT and loss; counts only, kept 7 days (168 hours).
      • The node journal: the warnings and errors of each node's agent and engines for 7 days, without client addresses, for every admin and in MCP.
      • The node doctor: host checks and safe fixes, among them Refresh the service, which rewrites the agent's systemd unit from the admin, a warning when the server's automatic security updates are off, and the fix Turn on security updates on Debian and Ubuntu.
      • Reachability, a section of its own with each node's week: probes inside Russia through Globalping, cheburcheck's volunteer probes and bsbord's mobile SIM probes (Crimea included, with a monthly cap), the antifilter block lists, a TSPU probe from your own node in Russia and the reports of kl!ck users from their own networks, each switched on separately and off by default, with an alert when a node is unreachable from the country; each node is also measured over its IPv6 address when its IPv6 works.
      • Response time through an open tunnel and connect time apart from it, and a notice when the panel's host loses UDP.
      • Alerts in the admin and in Telegram.
      • The panel watcher: a Cloudflare Worker checks the panel every minute and writes to Telegram when it does not answer.
      • Closing an alert until its problem goes away, and the doctor's 21 checks, among them CPU steal, connection tracking and the agent's open files.
    • Updates#

      • What's new: the release notes between your version and the available one, in the interface language.
      • Signed node agent updates, one node at a time, now or on a schedule, with a health gate and automatic rollback.
      • Panel self-update from signed GitHub releases, with a rollback when the new panel does not stay up.
      • Release channels (stable, testing, development) for the panel and its nodes, and a backup offered before the panel updates itself.
      • A one-line panel install that verifies the release signature.
      • Reproducible releases signed offline from the tag.
      • A download window for slow links, 30 minutes by default, and each node's download progress in the rollout.
    • Integrations#

      • API tokens with the readonly, operator and admin profiles.
      • The MCP server with plan / apply and owner approvals, now with day-to-day changes, one-call reads, probes between your nodes, diagnosis and scheduled tasks; ready prompts in the AI agent guide.
      • Telegram alerts to the admins who link their chat: health, releases, failed backups, plans waiting for approval, sign-in security.
      • The assistant: a chat with a cloud model that prepares changes as cards, with follow-ups and scheduled tasks.
    • Settings#

      • The admin, the user page, the node install page, Telegram messages and these docs in English, Russian and Simplified Chinese, the admin and the user page also in Persian and Turkish; sign-in with a passkey or a password with an authenticator code, optional Cloudflare Turnstile, an audit log. See Security.
      • A code audit (v0.1.84) went over sign-in, subscriptions, the node link, MCP and the assistant; the gaps it found are closed. See Security.
      • A decoy site at the panel's address hides the admin from strangers.
      • Encrypted backups to Cloudflare R2 on a schedule, with retention, or to a file on the owner's computer; restore in the panel or on a new server with the installer.
      • Panel logs, retention, export and metrics for the owner, and an opt-in reveal animation (Beta).
  2. Now#

    v0.1.34 – v0.1.81

    What the latest releases added is being tested in production; fixes and corrections to the docs come out of it.

    • VLESS in production#

      REALITY since v0.1.34, the node's own site as the target since v0.1.35, TLS and a CDN since v0.1.36. VLESS has not been tried behind a real CDN yet: in the end-to-end test the node itself stands in for the CDN. See VLESS behind a CDN.

    • Devices by device ID in production#

      Existing panels run in Only count, so you see who is over the limit before you switch to Limit. See How to move from Only count to Limit.

    • Xray JSON and sing-box JSON in production#

      No rule hands out Xray JSON by default: add one for the apps you want, or try an app first with ?format=xray-json on its link. See Who gets what.

    • A decoy site on every node#

      The built-in site, the panel's site or an archive of its own for each node. Nodes created before v0.1.36 stay on the built-in page until you pick another option. See Your own decoy site.

    • The country check in production#

      Every source is off by default. Switch one on and watch the From Russia column before you rely on its alert: the Globalping probes are not the panel's own, and listed is a registry mark, not a measurement. See Reachability.

    • Relays in production#

      New in v0.1.65. The panel checks a relay from outside the country, so a freeze that only people inside see shows up in kl!ck reports and bsbord, not in the panel's own check. See Relays.

    • IPv6 entry in production#

      New in v0.1.66, with its card in the admin since v0.1.67. Only on a node whose IPv6 works at the hoster. Whether people inside a country actually get through over it shows first in kl!ck reports. See IPv6 entry.

    • Ads blocked on the device in production#

      New in v0.1.67. Happ takes the list from its own geo files, by default Loyalsoldier's, which adds about 195,000 domains: on an iPhone, check that Happ keeps running with the preset before you give it to many people. See Ads and trackers.

    • The security updates fix in production#

      New in v0.1.67. It installs a package on the server: try it on one node first and read the doctor's result there before you run it on the others. See Fixes.

    • Nodes per group in production#

      New in v0.1.68. Turning a node off for a group takes it away from everyone in it at once, even from a person who has that node ticked on their own card: a profile on it stops reaching them. See Users and groups.

    • MTProto in production#

      New in v0.1.69 and experimental. It has not been tried against real blocking yet: give the link to one or two people first and check that Telegram connects from their networks before you hand it out widely. See MTProto.

    • The assistant in production#

      New in v0.1.81. It sends the conversation and what its tools read to the provider you choose, so read where each provider's data goes before you connect one, and set a monthly budget. Start with Only me under Who may use it, and read a plan card's wording before you press Apply. See Providers and where your data goes.

    • Blocking diagnosis in production#

      New in v0.1.81. block_diagnosis and access_report apply fixed rules to the nodes' counters and the country checks; where two causes look alike they say so, with low confidence. A lab of staged blocking checks seven of the classes before each release that touches them (an address, a port, UDP, a TCP freeze, a whitelist, one operator, a broken node); protocol fingerprinting and regional are tested on constructed counters only. The handshake counters need the agent v0.1.81. Read the evidence they list before you move people. See Assistant.

  3. Next#

    The next items in the queue.

    • Mistgate client#

      Our own desktop app, Windows first. It talks to the panel natively, not only through a subscription link, speaks Hysteria2 and AmneziaWG itself, has built-in detectors of TSPU and other blocking, and sends debug reports to the panel with the person's consent. The panel side is ready and off by default: the owner switches it on in Subscriptions → Who gets what, and the app will be able to enrol a device by a key and fetch servers directly; nothing changes for existing apps. The app itself is not released yet, and until then people connect with existing apps: see Client apps.

    • The servers' operating system from the panel#

      In the next release the panel installs security updates on one node and reboots it, with your confirmation: the node checks that a reboot is safe, the panel waits for it to come back and checks its profiles, and a failure stops the update and raises an alert. Several nodes at once with a canary, a time window and the panel's own server come next. Refreshing the package lists is the owner's click in the admin: API tokens and MCP cannot do it yet. See Update one server's system.

    • Update from another node#

      A node with a poor link to the panel gets the update file from a neighbouring node of the fleet that already has it. For now there is one fallback source, GitHub. See Download sources.

    • Which operator a connection comes from#

      The network (ASN) of each connection, counted per hour, never the address: an aggregate per operator, on by default, and a mode per person, off by default, both kept 7 days. It lets the diagnosis tell operators apart. Today the panel learns operators only from the country check's probes and the reports of kl!ck users. See Assistant.

    • A new IPv6 address in one click#

      The IPv6 entry and the reachability checks over IPv6 are in the releases since v0.1.65. The node agent already keeps an extra address on the server's uplink; the panel's half is missing: replacing the entry's address with a new one from the server's own /64 in one click. See IPv6 entry.

    • A relay that picks its best exit#

      Today a relay forwards to one exit server, and the app chooses between whole paths (direct, or through each relay) with the Auto group. Next, a relay gets a group of exits and the panel moves it to the one with the best link, from the loss and delay it already measures between nodes: not more often than a set interval, with the reason and the history on the relay's page, and with a manual pin. See Relays. Requested in #47.

    • Where your people are#

      A region in the first-run setup and in Settings: DNS presets, "direct" lists, recommended apps and reachability checks follow the countries you choose. China's pieces are already in (Chinese sites direct since v0.1.64; AliDNS, DNSPod and a Baidu check in the node doctor since v0.1.78); the setting itself is not.

    • Decoys that look like an ordinary self-hosted service#

      Built-in decoy sites in the style of a status page, a file share or a service's sign-in page, different on every installation, down to the 404, 429 and robots.txt answers, so that probers cannot recognise them; a sign-in there always fails and keeps nothing.

    • The Cloudflare edition on par with the VPS#

      Bring to the six-Worker edition what it does not have yet:

      • encrypted backups and restore;
      • MCP, the assistant, scheduled tasks and new API tokens;
      • probes, the debug capture, panel logs and the Load view;
      • client-eye tunnel checks;
      • updating the panel from the admin;
      • live node logs and an Own site decoy for a node.
    • Install from cloud-init#

      An unattended install of the panel and of a node from a cloud server's user-data, with ready examples for common clouds.

  4. Later#

    Planned, but not next in the queue.

    • Community routing lists#

      Ready rule sets maintained by the community (for Russia: a full list of local sites for "direct", and a mode for mobile whitelists), updated by themselves.

    • Any geosite or geoip category in a split rule#

      The split-rule editor suggests every category of the lists, not only countries.

    • A choice of the protocol core per node#

      The protocol cores (Xray, Hysteria2) are built into the node agent, so today a node runs the core its agent was built with. If a core ever changes in a way old client apps cannot follow and that cannot be bridged on the server, the panel will offer two agent builds, one with the older core and one with the newer, and let you pick one per node, with the same signed rollout and rollback. Until then a new core ships with the compatibility old clients need.

    • More protocols#

      Candidates: TUIC, NaiveProxy and MASQUE. Each plugs into the same subscription link through the protocol contract.

    • A Telegram bot beyond alerts#

      Today the bot only sends alerts and links an admin's chat. Later: running the fleet from Telegram, and a bot for users. See Telegram alerts.

    • Subscription mirrors#

      Copies of the subscription at other addresses, so that apps still get it when the panel's address is unreachable.

    • A shared map of blocks#

      Panels whose owners opt in send short anonymous reports of what fails (country, big operator, kind of protocol, what failed), and the project publishes weekly open summaries anyone can mirror. No addresses, names, ports or people ever leave the panel; reports show failures only and are labelled as unverified. Off by default. Not being built yet: first a local preview of what would be sent.

    • Moving between the VPS and the Cloudflare edition#

      Moving users, nodes and subscription links either way, with the links still working. Today a fleet is set up again for the other edition. See Moving.

    • The Cloudflare edition on a real account#

      Preview

      The first run of mistgate edge deploy on a real Cloudflare account: the six Workers, the router with the domain, and --rollback. So far the command and the panel have been tested against a simulated Cloudflare only. See Install the Cloudflare edition.

  5. Road to 1.0#

    The conditions for 1.0. Until then, breaking changes are possible.

    • The node protocol and the stored settings are fixed#

      A panel and agents of different 1.x versions understand each other, and the panel's data moves from version to version without manual steps.

    • Upgrades from older versions are tested#

      Every release passes an automated upgrade test from earlier versions: a panel with its data and the nodes connected to it.

    • A stable API#

      The API, API tokens and MCP tools without breaking changes within 1.x. See API and MCP.

Known limits#

  • Client-eye checks skip VLESS when the node agent does not report the required capability. Agents before v0.1.34 have no VLESS engine; TLS profiles also require vless-tls/1 (v0.1.36 or later).
  • Check ports looks from one vantage point: another node of the fleet or the panel host. On VLESS ports it runs a TLS handshake with the expected name, not a REALITY handshake.
  • The SSH installation logs in with the panel's own key, a password, or once with the owner's own private key, and the node card manages authorized_keys. Closing password login in sshd and rotating the panel's key are not there yet. See Install a node over SSH.
  • Backups go to Cloudflare R2 or to a file on the owner's computer. Restore is in the VPS edition (the panel and the installer); the Cloudflare edition keeps its data in D1 and has no restore yet.
  • The panel does not create a Cloudflare account, a bucket or a token: you supply a bucket-scoped token with read, write and delete access.
  • The Cloudflare edition is a preview, not yet tried on a real Cloudflare account: some features are not available in it yet (see the list), and a fleet cannot be moved between it and a VPS panel.
  • A device ID is what the app reports. Blocking refuses that ID; an app that reports another ID is a new device and takes a free place. To cut off a link that was passed on, issue a new link as well.
  • Speed per person is per node, so a person on two nodes at once can get up to twice the limit. Node agents before v0.1.39 hold each device to the smaller of the two limits.
  • In the Cloudflare edition the digest of what was sent to a node must fit one D1 row (2 MB), about 38,000 credentials per node. See the limits of this edition.
  • sing-box JSON has no XHTTP and cannot pin a self-signed Hysteria2 certificate; Gecko and AmneziaWG go out only in Mihomo YAML. A server a format cannot carry is left out of that format. See What works where.
  • The country check looks from a few probes, not from every network, and the block lists cover Russia only. listed means the address is in a registry, not that its traffic is cut.
  • Some defaults are country-specific (Yandex DNS for Russia, AliDNS and DNSPod for China, the doctor's control domains, split-DNS presets); all of them are settings.
  • The assistant sends the conversation, its instructions, your notes and what its tools read to the cloud provider you connect, so that provider's data policy applies. Prices you enter are estimates, and the monthly budget is measured from them. See Providers and where your data goes.
  • The assistant, MCP, scheduled tasks, probes, the speed test, the debug capture, panel logs, panel resources and the Load view are in the VPS edition only. See the list.
  • The diagnosis is only as good as the counters. VLESS keeps no attempt counter, and the kernel AmneziaWG backend cannot tell a handshake initiation from an answer; where two causes look alike, block_diagnosis says so with low confidence. Agents before v0.1.81 do not count why handshakes fail or what a node's machine is short of: the panel says "not known".
  • Scheduled tasks follow the panel's fixed UTC offset and do not follow daylight-saving time: set the offset again when local time changes.

Edit this page on GitHub