mistgate Docs
EN
300

In the admin: Nodes

Nodes

What the Nodes list and a node's page show, what each status means, and how to change, restart or retire a node.

On this page

A node is a server that runs the Mistgate agent, mistgate-node. The panel tells the agent which profiles to run, and the agent reports back its state, its traffic and what it sees on the host. This page walks through the Nodes section and a node's page tab by tab. To connect a new server, see Add a node.

The Nodes list#

The header says how many nodes there are, how many people are online and how many nodes have problems ("3 nodes · 12 online · 1 with problems"). Add node opens the dialog that installs a node over SSH or makes a manual install command: see Add a node.

Above the table:

  • the status chips All, Problems and Healthy, each with its count;
  • All protocols: shows only the nodes that run a profile of one protocol;
  • All countries: shows only the nodes of one country.
Column What it shows
Node The status dot, the name, the location (or country) and the provider. Under it, the reason of the status, or the status word, and the first two lines of the node's notes (the full text on hover).
Address The domain or IP clients connect to.
Protocols The protocols of the profiles on the node.
Online People connected now.
Today Traffic since 00:00 UTC.
CPU The latest CPU reading.
Uptime Host uptime as the agent reports it.

On a phone the table becomes one card per node with the same facts. The list refreshes every 10 seconds. Retired nodes are not listed.

List, Map and Load above the table switch the view, and the browser remembers the last one. A phone has no map; the Cloudflare edition has no load (see Load).

A healthy node whose agent is older than the version this panel ships reads Healthy · outdated agent with its version; the line is a link to Updates. See Updates.

Bulk actions#

The owner can tick nodes in the list (on a phone, on each card) and act on all of them at once. The box in the table's header ticks every shown node; Select ▾ next to the status chips replaces the selection with All shown, With problems, Password only or Outdated agent, or clears it. At most 50 nodes are selected at once, and a retired node cannot be.

While nodes are selected, a bar floats at the foot of the page (on a phone it takes the place of the menu dock): how many are selected and how many of them the filter hides, then the actions. Esc clears the selection.

Action Where What it does
Ports Check ▾ Runs the port check on each node, one node at a time (a check takes about 10 seconds). Starts at once.
Doctor Check ▾ Asks each node for a fresh doctor report, four at a time. Starts at once.
Update agent the bar Opens the rollout window of Updates with the selected nodes whose agent is outdated; the rollout then runs on Updates. Off while another rollout runs.
Change SSH password the bar Asks first. The panel then generates a password of its own for each node, changes it on the server and checks that it logs in. The passwords are not listed: reveal each one in its node's Server → SSH access card. The old passwords stop working.
Panel key More ▾ Puts the panel key on the nodes that log in with a password only, as the offer above the list does. Starts at once.
Bind own key… More ▾ Adds one public key line to ~/.ssh/authorized_keys of each node. The window shows the key's type, comment and fingerprint before anything is sent; a private key or several lines are refused in the browser.
Unbind a key… More ▾ Removes a key by its public line or its SHA256: fingerprint. The panel key is not unbound this way: do it in a node's card. A node that did not have the key reads Was not there.
Decoy site… More ▾ Sets one decoy site option on all of them in one change. See Decoy site.

Not every selected node takes part in every action, and the window and the progress say which ones sit out and why: an offline node gets no check, a node that logs in by key only has no password to change, a node with no saved SSH access gets no SSH action. The SSH actions read the nodes' saved access fresh before they start, ask you to confirm your sign-in once, and go one node at a time: the panel logs in to each server several times.

A running action grows the bar upwards into a row per node: Waiting, Running…, Done (a check says what it found, with Open to the node's page), Failed with the reason, or Skipped with the reason. Stop lets the node in progress finish and starts no more; the rest read Not started. At the end a summary says how many went through: Retry for N runs the failed and the not started ones again, Close puts the rows away and keeps the selection. One action runs at a time, and leaving the page while it runs asks first. If the servers stop answering after several retries, wait a minute: each server limits how often the panel logs in.

Load#

Load reads each node's hours of host load over the last 24 hours, 7 days (the default) or 30 days, and says how loaded the node was, what runs out first and how many more people it can take. It never contacts a node. The VPS edition only: the Cloudflare edition does not keep these hours.

Above the table: the nodes by status, the room left across the fleet ("room for about 80 more people") and the node with the most room.

Column What it shows
People at peak The most people online at once in the period.
Busy-hour CPU The CPU mean of the busiest hour: the hour the node has to get through every day.
Runs out first The resource closest to its limit once it is halfway there: CPU, CPU steal, bandwidth, memory, connection tracking or disk; Nothing close otherwise.
Room for How many more people the node can take before its busy hour reaches 80% CPU (and, with a known bandwidth, before its busier direction reaches 90% of it). No room on an overloaded node. A dash when the data cannot tell, with the reason on hover: under a day of readings, a number of people online that barely changed, or a CPU that does not follow the people.
Status OK, Busy (worth a look), Overloaded (people already feel it) or No data.

A node is Busy from a 60% busy hour, typical CPU peaks of 85% (the 75th percentile of the hourly peaks: a few moments at 100% in a day, such as the server's own updates, do not count), 10% steal, 70% of its bandwidth, 85% memory, 80% connection tracking or an 80% full disk, and Overloaded from an 80% busy hour, 25% steal, 90%, 95%, 95% and 92% (CPU peaks alone never make a node overloaded); the steal, connection tracking and disk thresholds are the doctor's. The bandwidth is the one set in the node's settings, else the link speed its network card reports.

The arrow at the end of a row opens what to do and the numbers behind it. The recommendations name the node with the most room to move people to, say when the provider oversells the host (high steal), when connection tracking, memory or the disk fills, when the server swaps, and when the agent's memory grows day after day while the people do not (a likely leak). The cost per person, "CPU per 100 people" and "traffic per person", comes from a least-squares fit of the hourly CPU and traffic on the hourly mean of people online. An agent older than this panel does not report steal, swap, connection tracking or its own cost: those say "not known", never zero, and the node asks for an update. The assistant and MCP (capacity_report) read the same report.

Node statuses#

Status What it means What to do
Waiting for install The install command was made, the agent has not connected yet. The line says how long the command still works, or that it has expired. Run the command on the server, or make a new one (New install command on the node's page).
Healthy The agent is connected and nothing is wrong. Nothing.
Needs attention The agent is connected, but: there is no enabled profile on the node (users do not get it), the node's state differs from the panel's, or the server clock differs from the panel's. Read the line under the name. Add a profile, restart the profiles, or fix the clock.
Partly working Some profiles failed to start, others run. Open Profiles on the node and read the error under each failed profile.
Broken Every enabled profile failed to start, or a host check of the doctor failed. Open Profiles or Doctor.
Unreachable No link to the agent, and the host did not come back within 10 minutes. Check the server at the hoster, then restart the agent over SSH (the banner gives the command).
Host blip The link dropped less than 10 minutes ago, or came back with the same boot time. It is not counted as a problem. Usually nothing: short drops are the hoster.
No traffic The agent is connected, but the client-eye checks fail: clients do not get through. The banner says why when it knows (a blocked UDP port, a certificate, the exit). Restart the profiles or change the port.
Updating The agent is being updated. Wait. If the check after the update fails, it rolls back by itself.
Retired The node was retired from the fleet. Nothing.

Problems are the red statuses and Needs attention, except the clock warning: that one is shown but not counted. These counts appear in the header of the list, on the Overview and in the menu.

A node's page#

The header shows the country code, the name, the status, and a WARP chip when the node has a WARP account (the chip reads the Cloudflare data centre while WARP works and links to the WARP card). Under the name: the location, the provider, the address, the agent version, and outdated agent when there is a newer one.

Two buttons sit on the right:

  • Restart profiles restarts every profile of the node. The dialog says how many people drop for a few seconds; they reconnect by themselves. The button is off while the agent is not connected or the node has no profiles.
  • Doctor opens the Doctor tab.

Under the header a banner explains the current state with the action that helps:

State Banner
Unreachable "The node has not answered for …", the time it was last heard from, and a command to copy.
No traffic "Alive, but no traffic" with the reason from the health checks, Restart profiles and Check now.
Host blip "The host blipped" with Got it.
Updating "The agent is updating".
Waiting for install "Waiting for the agent" with New install command.
Retired "This node is retired".
No profiles "No profiles on … yet — users will not get it" with Add profile.
A profile failed The profile name, the cause in words, and Open profiles.

The command of the "Unreachable" banner, to run on the node over SSH:

Shell
systemctl restart mistgate-node && journalctl -u mistgate-node -n 50 --no-pager

The page has seven tabs, each answering one question: Overview (is all well?), Ways in (how do people reach this node?), Profiles (what runs here?), Exit (where does people's traffic leave?), History (what happened?), Doctor (what is wrong on the host?) and Server (what is this machine?). The Ways in tab carries a red badge with the number of lanes broken on our side (a relay that is down, a port that does not answer); a block by a country does not count, that is an alert and the amber colour of the lane. The Profiles tab shows how many profiles failed to start, the Doctor tab how many items need attention. A retired node keeps Overview, History and Server. Links and bookmarks made for the tabs of earlier versions (Users, Logs, Journal, Events, Settings) open the place those moved to.

Overview#

  • CPU, RAM, DISK, NETWORK: the latest readings. The CPU bar shows the softirq part (network interrupts) in a second colour, with the vCPU count and the load. Readings arrive a few seconds after the agent connects; an offline node keeps its last readings and says since when it is offline.
  • Resources: what took the machine's CPU, split into Mistgate (the agent and every engine inside it: what Mistgate itself costs the server), Kernel networking (softirq), Steal (time the hypervisor gave to other virtual machines) and Everything else; memory with Mistgate's share and swap; the connection-tracking table and the agent's open files as meters (amber from 80%, red from 95%); sockets; and the link against the speed of the network card when its driver reports one. Each part explains itself on hover. An agent older than this panel does not report these: the card asks to update it.
  • Load history: the node's hours over the last 24 hours or 7 days, the mean and the peak of each hour: CPU, people online, and steal and Mistgate's own CPU once an updated agent has reported them. The VPS edition only.
  • Traffic · 24 h: hourly traffic through the node.
  • Ways in: a one-line strip with a chip per lane; a chip opens that lane on the Ways in tab, and All ways in › opens the tab.
  • Alerts: the open alerts of this node, with the same buttons as on the Health page, and the last three that closed. All alerts opens the Health page. The card is hidden when there was nothing.
  • Client-eye checks: one row per profile with its latency and exit IP, or why it fails. A click on a row opens its 24-hour history. Check now starts a round for this node.
  • On this node now: how many people are connected, per protocol, and who they are, with the device model, the protocol and the current download speed. Six people show; All N opens the rest in place.
  • Top today: the people with the most traffic through this node since 00:00 UTC, five at first; All 10 shows the ten.

The checks, alerts and the doctor are described in Health.

Ways in#

This tab shows every way people reach the node as one "metro line". On top is a chip People · Russia (the country of the country check), under it a vertical line with one branch per lane, and at the bottom the node itself. The lanes:

  • Direct: the node's own address and the number of ports, with a line under it: "From Russia: all networks", "Dark on MTS, Tele2", "Does not open from Russia" or "From Russia: no data".
  • Through a CDN: the node's VLESS profiles behind a CDN, if it has any, and their number. People connect to the CDN, not to the node, so the line under it says the country sees the CDN, not the node.
  • Through msk1: one lane per entry node that relays this node, with the number of servers and a line saying whether people reach the entry on the networks where this node is dark ("msk1 is seen from MTS"). Only the owner sees these lanes.
  • IPv6 entry: an extra server for people whose network blocks the node's IPv4 address; off until the owner turns it on (see IPv6 entry). The lane is shown when the server has IPv6 or the entry is on. It reads "off", the entry's address when it is on, or "not available" with the reason under it. It opens the entry's card.
  • Add a way in…: a dashed row, for the owner only. When Direct is dark on some networks it turns amber: "Dark on MTS — add a way around". With no other ways yet it says: "If people's network does not see this node, give them another way in: through another node or over IPv6."

On an entry node, below the line, the group Reached through msk1 has one row per exit ("→ DE de1 · 2 servers"), then Relay another node… and a footer: relayed traffic crosses the server twice, so it counts twice in the server's traffic and load.

Each lane has a ring: green, works and every network sees it; half amber, works but part of the country does not see it; red, down or blocked; grey and hollow, off. A click on a lane opens its detail on the right; on a phone the list and the detail take turns, and ‹ All ways in goes back. The open lane is kept in the address (?tab=ways&way=direct, ipv6, via:<entry id> or out:<exit id>), so a link or a reload opens it again; without it the first lane with a problem opens, else Direct. The arrow keys, Home and End walk the list. The old link ?tab=profiles#ports opens Ways in → Direct.

Direct#

"People connect to 203.0.113.7." Under it come the Ports block, then Is de1 seen from Russia: the country check row and, when the block is partial, the breakdown by network. Buttons: Check ports and a link Profiles ›. With no profiles it says "No profiles: people have nothing to open." and offers Add profile. While the country check has nothing on the node it says "The country check has nothing on this node yet." and, for the owner, links Sources on Reachability.

In subscriptions (for the owner) lists the node's servers on its own address, with a switch each. Off, people do not get the server at the node's address, while it keeps running: its servers through other nodes and over IPv6 stay, and relays forward to it as before. Use it when people should reach a node only through another node, say a node abroad that people reach through an entry in Russia. When turning one off would leave people no way to the server (no working relay, no IPv6 entry, or AmneziaWG, which cannot go through another node yet), the panel asks first and says why. When every server is off, the lane reads Not in subscriptions. The Through a CDN lane has the same switches for the profiles behind a CDN.

Ports#

The Ports block sits in the Direct lane of the Ways in tab (it used to be on the Profiles tab). Check ports runs two checks from one connected fleet node (or the panel host when available): UDP delivery and TCP reach. Each row shows the port, the verdict, when and from where it was checked, and the profile that listens there.

UDP. The check sends probe packets from one connected fleet node (or the panel host when available) to this node and asks its agent how many arrived. It checks delivery from that sender's network path. A passing result does not prove that every client can reach the port, and a failure can depend on the sender, route or source-dependent filtering.

When a port loses packets, the panel sends again at once from a second node on the other side: a node outside the CIS when the first was in it (Russia, Belarus, Kazakhstan, Kyrgyzstan, Armenia, Azerbaijan, Tajikistan, Uzbekistan, Moldova, Turkmenistan), one inside when it was not, then another provider or country. The row shows both measurements, "from minsk1 118/150 · from de1 150/150", and who is at fault:

  • Lost from both sides: the node's port loses packets. The port counts as lossy and raises port_lossy.
  • Lost from one side only: the loss is on the way from that sender, not on this node. The port counts as fine; the sender gets the alert path_lossy and is asked last for the next 24 hours, so one weak path does not colour every check. When the ports that sender lost got nothing at all (at most 1 %) while its other ports arrived whole (often only 443), the path is filtered rather than lossy, typically the hoster's DDoS protection against data-centre traffic; the alert, the event and a line under the UDP rows say so. The node is fine and people on home networks usually connect, but port checks (and a relay) from that sender cannot pass.

If both senders get at least 95% of UDP 443 but at most 1% on every other tested port, Mistgate marks the node as filtered from check senders. It stops recruiting senders and confirmation runs, rechecks it at most once a day with a sender already used, and clears the mark after a clean run. A panel protocol probe through that inbound that passed after a port check, or at most 15 minutes before it, suppresses port_lossy for it. Such a row reads Test blocked in amber instead of the loss, with a line naming the protocol whose check passes, and the Direct lane counts the port as working. A node with the mark gets a line under the UDP rows in place of the sender's: its hoster lets only UDP 443 through from our nodes, so port checks say little there.

A port is also lossy when it delivered clearly less than the other ports of the same run (the median port), by more than 4 % of the packets and more than the usual spread on a path that loses a little everywhere.

TCP. The check opens a real connection to each TCP port the node should answer and asks whether it reached Mistgate itself, not another server in front of it. Port 80 is checked when the node has an ACME certificate name: the agent arms a one-time token and the sender requests it the way Let's Encrypt does. The Hysteria2 site port and each VLESS TCP port get a TLS handshake for the expected name. A VLESS profile that fronts another site's REALITY target is judged by that target's certificate.

Verdict Meaning
Reaches Mistgate The connection got to Mistgate on this node.
Another server answers Something else answered: a provider filter that replies, a proxy or a web server in front of the node.
Held by another program on the node Port 80 only: another program on the node already listens there, so Mistgate cannot answer on it.
Closed The connection was refused.
No answer (filtered) The connection timed out, which usually means a firewall that drops it.

A row with Not checked and a short reason could not be probed (the agent is too old, the sender has no route to the node, another check is running, or it failed). When no sender can probe at all, the last stored result stays. A sender on the node's own host proves nothing about the provider's filter, so the panel says so and saves nothing.

Under the rows, one line says whether the node's Let's Encrypt certificates renew:

  • Port 80 is open: they renew over port 80.
  • Port 80 is closed, but TCP 443 answers Let's Encrypt: they renew over TLS on 443. This needs a listener that answers on 443, such as the Hysteria2 site or a VLESS profile whose REALITY target is This node's site.
  • Port 80 is closed and TCP 443 does not answer Let's Encrypt: the certificates will stop renewing. The line names the VLESS profile that holds 443 with another site, if there is one. Open TCP 80 at the provider, move VLESS to 8443, or make its REALITY target This node's site.

There is no line for a node without an ACME certificate name, or when port 80 could not be probed. The same port 80 result decides the certificate sentence in the warning you get when a VLESS profile pauses a Hysteria2 site: VLESS and Hysteria2 on one node.

On VPS installations, the panel repeats the check every six hours for each online node with the udpcheck capability and an enabled inbound. It immediately confirms a lossy or broken result with one more check; the alert follows the latest verdict. The TCP rows are refreshed by the same scheduled run; they raise no alert of their own.

Through another node#

The lane Through msk1 says: "People connect to msk1; it passes their traffic to de1 unchanged. The keys stay the same." The block Servers holds the relay rows: the profile, the port on the entry, where it forwards to, the state, the two checks (see Relays), the buttons Turn on, Turn off, Check now and Remove, and the supervision line (Supervision). Next comes Is msk1 seen from Russia. Buttons: Check all, Add server… and a link Open msk1 ›.

On the entry node the lane → de1 is the same: the relay rows for that exit, Check all, Add server… and Open de1 ›.

IPv6 entry#

"An extra way in for people whose network blocks this node's IPv4 address. Not the same as the IPv6 exit for people's traffic." (see Exit).

What it does. With the entry on, every person who may use the node gets one more server, "<name> · IPv6", next to each server of the node, in the list their app shows ("🇩🇪 DE · Hysteria2 · IPv6"). It connects to one of the server's own IPv6 addresses (by default the one the server uses for its outbound IPv6) and is otherwise the same server: the TLS name stays the node's domain, REALITY keeps its target and keys, a self-signed certificate keeps its pin, and the person's credentials do not change. Apps without IPv6 skip it. The main address, its DNS records and the certificates are not touched. People get the new servers at the next refresh of their subscription (every 12 hours by default, see Subscriptions). The servers of a relay stay as they are and are not doubled for IPv6: the order is the node's server, its IPv6 server, then the relay's servers.

What it does not cover. A server that people reach through a CDN (they connect to the CDN's address), a server whose certificate is issued for the node's IPv4 address, and the static AmneziaWG keys (.conf file and vpn:// key): a key keeps the address it was issued with and does not update, so changing it would break it without a word. The Mihomo proxies of AmneziaWG refresh with the subscription and do get the IPv6 server (see AmneziaWG). AmneziaWG over IPv6 carries 20 bytes more of header per packet: a profile whose MTU is set tight for IPv4 may fragment there; the default of 1420 leaves room.

Who gets it. Everyone who may use the node. The panel does not know the network of a person, and an app without IPv6 simply does not choose the server, so there is nothing to choose for.

When it is available. The panel needs the node's Doctor to report a working IPv6 uplink. When it cannot be turned on, the panel says why (on the card and, in MCP, in node_get): no_ipv6 (the server has no global IPv6 address), ipv6_broken (it has one, but outbound IPv6 does not work at the hoster, so no connection would come in either), agent_old (IPv6 works, but the node agent does not list the server's stable addresses: update it), not_checked (the doctor has not reported yet) or main_is_ipv6 (the node's address is an IPv6 address already). The address is one of those the doctor lists for the server's uplink, by default the one it uses for outbound IPv6 when that is one of them (a temporary address never is); there is no free text, because an address the host does not hold fails the check anyway.

Turning it on is checked first. The panel sends the same packets as Check ports to the IPv6 address, from outside: UDP datagrams to each of the node's UDP servers (counted at the node) and a TLS handshake to each of its TCP servers, where the node's own certificate, pin or REALITY chain proves that it is this node that answered. Servers that get no IPv6 server (behind a CDN, or with a certificate for the IPv4 address) are left out, and so is port 80: Let's Encrypt validates the main name over IPv4. The sender is another node of the fleet whose IPv6 works (two at most, in the order of Check ports); the panel sends itself only when no node can and it has IPv6 of its own and is not on the node's host. Only when neither another node nor the panel can send at all, Globalping (when it is on) connects to one TLS port from two probes anywhere: that shows that the address answers a connection, not that it is this node, and the result says verified by Globalping. When nobody can check, the entry is not turned on (no_ipv6_verifier): an address nobody could test is never passed. A sender that cannot run its part fails the check too, and so does a part that nobody can check while the other one could be. A passing check that is less than 10 minutes old is used; otherwise turning the entry on runs one. A port that does not answer as the node's own fails the check and nothing changes. The check writes nothing to the Ports results of the main address.

Turning it off takes the servers out of the next subscription; people connected through one pick another server. It always works, also when the node's IPv6 has broken.

Alerts. While the entry is on, the IPv6 rounds of Globalping measure the entry's address, and a block over IPv6 raises an alert of its own for the node, from the rounds since the entry was turned on (see IPv6); without the entry IPv6 is information only.

The card. The lane opens the card IPv6 entry on the node's Ways in tab, with the subtitle above. It has these states:

  • Off. "Off. Subscriptions connect to {address} only." with the button Add IPv6 entry…. Under it, when the country check knows, one of two lines: "IPv4 is blocked from Russia, IPv6 gets through." or "IPv4 is not blocked from Russia now: this would be a spare way in."
  • The dialog. "IPv6 entry on de1": "People get one more server, “de1 · IPv6”, next to each of this node's servers. Apps without IPv6 skip it; nothing changes for the main address." Below it is a list of the node's uplink addresses to choose from; the one the server sends from is chosen first and marked "(the server's address)". On a node with AmneziaWG the dialog adds: "Static AmneziaWG keys (.conf, vpn://) get no IPv6 server: they keep connecting to {address}." Only check checks the address and changes nothing. Check and turn on uses a passing check that is less than 10 minutes old, or runs one first ("Checking {addr} from outside…", about 15 seconds). A failed check changes nothing.
  • A failed check. "{addr} is not reachable from outside: tcp/443 no answer (filtered). Nothing changed.", with the verdict of every port as a small chip. When nothing could check it: "{addr} was not checked." and the reason, for example "Nothing can check this address: no other node has working IPv6, and Globalping is off or could not test it."
  • On. "On: {addr}. Checked {ago} from {sender}. Subscriptions show “de1 · IPv6”." When only Globalping could check it, the card adds: "Only Globalping could check it: a connection to the address opened, which shows it answers, not that this node answers." Turn off asks first: "“de1 · IPv6” leaves the subscriptions; apps drop it at their next refresh." If the server's IPv6 stops working while the entry is on, the lane shows Down and the card asks to turn the entry off.
  • Not available. There is no button; the card shows the reason: the server has no IPv6, IPv6 is broken at the hoster, "Update the node agent to use an IPv6 entry.", "The main address is already IPv6." or the doctor has not reported yet.

The last check, passed or failed, stays on the card, so the result of Only check is still there after a reload.

Only the owner can check, turn on and turn off the entry; a helper and a read-only admin see the state without buttons. The API refuses tokens. The same can be done through MCP: node_ipv6_check (a read), and node_ipv6_plan / node_ipv6_apply, which wait for the owner's approval (MCP). The audit log records node.ipv6_check and node.ipv6_entry (on or off, with the address).

Add a way in#

For the owner. Add a way in… opens the dialog "A new way in to de1" (?tab=ways&add=1):

  1. When the node is dark it starts: "From Russia, de1 does not open on MTS. Give those people another way in — the node and its keys stay the same." Choose Through another node ("People connect to a node their network sees; it passes their traffic to de1 unchanged.", with the best hint: "Good fit: msk1, seen from MTS") or IPv6 entry. It opens the entry's dialog when the node can take an entry and it is off; otherwise it stays greyed, with the reason or "Already on: {address}".
  2. Through which node: Good fits: seen from MTS come first, then Other nodes. Each shows its country code, whether people see it and, where the TSPU probe runs from it, whether it reaches this node. A node that already relays servers of this one says "Already passes on 2 servers of de1".
  3. The relay dialog, limited to this node's servers: tick the servers, the port on the entry (auto), Only check or Check and turn on, then the progress and the result per server. Back returns to step 2.

The mechanism is still called a relay in the dialogs and in Relays; "way in" is what you look for on the node's page.

Who sees what#

The owner sees every lane. A helper and a read-only admin see Direct and IPv6 entry, read-only: no relay lanes and no Add a way in….

Profiles#

Every profile on the node is a row. The name links to the profile; under it is the protocol.

For Hysteria2 the row shows:

  • Port: the UDP port on this node;
  • Domain (SNI): the server name in the certificate;
  • Certificate: the expiry date (amber with less than 14 days left, red with less than 3), self-signed, pinned, or none: the profile is not running.

For AmneziaWG the row shows Port, Backend (kernel or userspace, with its version), Egress (Direct or WARP) and Devices ("5 · 2 online"), and a status line with the last handshake.

The Port cell of each profile has a small dot with the last port check: works, loses packets or does not answer. It links to Ways in → Direct, where the Ports block and Check ports are.

Under each profile a 24 h line says what the node counted of connection attempts on it: connections, dropped at once (accepted, then frozen or cut within seconds: what the TSPU's freeze looks like), key refused (the key's term ended; a removed or rotated key is unknown to the node and counts among the unknown keys), for Hysteria2 the share of started handshakes that never authenticated (the panel's own checks are left out), for AmneziaWG how many devices never completed a handshake, and unknown keys (scanners send them too). With an agent from v0.1.79, VLESS profiles count unknown keys too: requests with a UUID the node does not know, from a removed device or an old subscription. With an agent from v0.1.57, Hysteria2 profiles also say how many connections froze (the node kept sending and nothing came back: how a TSPU freeze looks), the average RTT and the loss; early drops are then not marked as trouble, because most are short checks of the app. An agent that counts why handshakes fail (links/3) says that too: answered, no reply (the node answered and nothing came back), certificate refused (apps did not trust the node's certificate: our side), refused by the node, for VLESS REALITY wrong REALITY key and sent to the decoy, for VLESS froze after 16 KB (the TSPU's freeze of a TCP stream) and stalled, and for AmneziaWG on the userspace backend answers lost (the way back to the device is cut) and not answered by the node (our side). Hover a fact for a short explanation. The kernel AmneziaWG backend cannot tell an initiation from an answer, so it shows neither. Counts only, never an address. A node whose agent does not count connections yet says so once above the profiles: update the agent.

On the right is the state: Running, Starting, Failed to start or Off. A profile that failed shows the cause in words and the agent's own error text under it. When the cause is clear, a button fixes it: Change port (the dialog opens on a free port), Set the domain, or Open WARP.

Each row has:

  • Edit: the port and the Domain (SNI) for this node only, and the Enabled switch. A profile switched off stays on the node but leaves the subscriptions.
  • Restart: restarts only this profile. Its connections drop for a couple of seconds.
  • Remove: takes the profile off the node. The dialog says who notices: the users who lose the node at their next subscription update, or, for AmneziaWG, the devices that stop working. Their keys work again if you put the profile back on the same node. If relays pass the profile's servers, the dialog also says how many go with it.

Add profile opens the dialog that puts a profile on this node: see Profiles. When the panel has no profile at all, the dialog offers to create a Hysteria2 profile on port 443 (or the first free port) and put it on the node in one go, with Set up by hand as the other way.

Exit#

Where people's traffic leaves the node. Top to bottom: the WARP card, People's traffic and DNS choice on the user page (on a wide screen it stands on the right). WARP is described in WARP; the WARP chip in the node's header and every Open WARP in the panel lead to its card.

People's traffic holds three settings with one Save, which sends only what changed; changing the torrent setting restarts the node's Hysteria2 profiles.

IPv6 exit (IPv6 for clients in earlier versions) is on by default (the way every node worked before), and only the owner changes it. It decides whether people's traffic may leave the server over IPv6; it does not change how they connect to the node. Turn it off when the provider's IPv6 addresses are located in a different country than the node: an app picks IPv4 or IPv6 per site, so people would appear to come from two countries at random (over IPv4 the exit reads as the node's country, over IPv6 as the other one, and Google or YouTube then show the wrong region). With it off, the node answers client IPv6 that would leave through its own uplink at once instead of dropping it silently: a TCP connection gets a reset, anything else an ICMPv6 "administratively prohibited", so apps move to IPv4 instead of waiting for a timeout. A kernel without the nftables reject action cannot do that: the node then drops that IPv6 silently (apps may wait for a timeout; if even that rule is refused, IPv6 is not blocked at all) and writes a warning event, and the tunnels keep working. The agent retries every 10 minutes until reject works again, then records a recovery event.

  • AmneziaWG. The agent adds one reject rule to its own firewall table (mistgate_awg) and removes it when you turn the setting back on. The ::/0 route stays in the tunnel, so a person's real IPv6 never leaves outside the VPN. Configs the panel issues from now on (the subscription, the app, the .conf and the vpn:// key) carry no IPv6 address of the tunnel either whenever the node gives clients no IPv6, that is when this setting is off or the node has no working IPv6 at the hoster or has not been checked yet: with no IPv6 address of its own the device refuses IPv6 at once and the app uses IPv4. A config issued earlier keeps its address and is helped by the node's reject rule alone. The node can only answer packets that reach its firewall: with no IPv6 route at all, the kernel itself answers "no route", which some systems treat as a soft error and wait out. IPv6 addressed to the tunnel's own address is untouched. IPv6 DNS servers of a preset are rejected too, so the device moves on to the IPv4 ones.
  • Hysteria2 and VLESS. The node's direct outbound connects over IPv4 only, and a name that has only an IPv6 address does not open. Connections that are already open keep the address family they have; new ones go over IPv4. The node's own name resolving (DNS resolvers for this node) works as before.
  • Exit through WARP is not touched. Its exit is Cloudflare's, not the node's uplink, so the geolocation of the node's range does not matter there.

The setting needs an agent that reports client-ipv6/1. An older agent shows the switch greyed out with "Update this node agent to turn this off", and the panel never sends it the setting; turning IPv6 back on always works. Each change is written to the audit log (node.update with client_ipv6 on or off).

Block recognized BitTorrent traffic is off by default. It blocks plaintext BitTorrent that the node's users start, on AmneziaWG in the kernel and on Hysteria2 in its engine, and writes a torrent_attempt event without any address. It needs an agent that supports it. See Torrent protection.

DNS resolvers for this node: Server's own resolver (the default), Yandex DNS or Cloudflare + Google; a list of your own (up to 8 IP addresses, ip:port works too) shows as Custom DNS and is edited there. The node and its VPN engines resolve names with them, the names in its users' traffic included. With Server's own resolver the node uses whatever the server uses. For a node in Russia pick Yandex DNS, so Russian services such as gosuslugi.ru resolve reliably; for a node in China, use Custom DNS with AliDNS and DNSPod (223.5.5.5, 119.29.29.29). They are not the DNS presets that apps receive, and the doctor checks the server's own resolver separately: see DNS.

DNS choice on the user page is a separate block of the node, not to be confused with the resolvers above. Here the owner ticks the DNS presets that people may pick for this server on their page (the order is the order on the page) and marks one as the default for the node. Only the owner changes it, and the change is written to the audit log (node_dns_options). A node with no presets offered gives no choice. It matters only when DNS choice on the page is on in Subscriptions → User page, and it reaches AmneziaWG keys and the AmneziaWG proxies of a Mihomo profile; see DNS.

History#

What happened on the node, in three parts switched at the top of the tab: Events, Journal and Logs. The part is kept in the address, so a link or a reload opens it again.

Events#

The history of the node, grouped by day: connected, went quiet, came back, profiles added, started, failed and restarted, settings applied, agent updates and rollbacks, certificate renewals, the AmneziaWG module build, and recognized BitTorrent attempts. Related events are joined into one line. A torrent event includes the inbound and the protocol, never an address (neither the client's nor the destination); the panel adds a user's name only when the node can identify that user reliably.

The chips All, Problems, Profiles and Agent filter the list. Details opens the raw events behind a line: the event code, the exact time, the source (node agent, panel or admin action) and the parameters. The log is paged: 50 events at a time (25 or 100 on request), newest first, Older goes back in time and Newer returns to the page you came from; the page is in the address, so a reload or a copied link keeps it. When a filter finds nothing on a page, Search further goes to the next, older one.

Journal#

The node's warnings and errors, saved by the panel for 7 days: what the agent and its engines (Hysteria2, VLESS, AmneziaWG, xray, WARP) wrote at the Warn and Error levels. Each row is one kind of record in one hour: the time of the latest, the level, the source, the profile when the record names one, how many times it came (×12), the message and its details. Client addresses are cut out on the node and again in the panel, so every admin can read the journal; the message is the template the engine logs, without the values of a single connection.

  • 24 hours or 7 days; Warnings and errors or Errors only.
  • Search the journal looks in the messages and the details on the panel, across all the rows of the period. At most 200 rows are shown, newest first; when there are more, the tab says so.

A node whose agent does not send the journal yet says so: update the agent. On the Cloudflare edition, where the live tail is not available, Logs leads here. The same journal is in MCP as node_journal.

Logs#

A live tail of the agent and its engines, relayed by the panel. The last 200 lines arrive first, then new lines as they are written; up to 1000 lines stay on screen.

  • All, Info, Warn, Errors filter by level; Search the log filters by text.
  • Pause freezes the screen while new lines keep arriving; Resume shows them.
  • Download saves what was received as <node>-agent.log.

This part works while the agent is connected (Healthy or No traffic). If the link drops, it reconnects every 5 seconds. Only the owner can read the logs: they contain client addresses. Every admin can read the Journal.

Doctor#

The node's own host checks, grouped into Needs attention, Accepted as normal, All good and Not checked, with the fixes the panel can apply and the manual steps for the rest. Run again asks the node for a fresh report. The first report comes about 30 seconds after the agent connects; an agent that is too old for the doctor gets a link to Updates. See Health.

Server#

What this machine is. On a wide screen, two columns: on the left one form (Node, Capacity and Agent link) with one Save at its end, and the Agent card under it; on the right SSH access (a node installed over SSH), Decoy site, AmneziaWG backend and Host. The Danger zone comes last. On a narrow screen it is one column in the same order.

  • Agent: the version, outdated agent with a link to Updates when it is behind, what an older agent cannot do (see Agent version and drift), and New install command (the owner only) for a lost certificate or a reinstalled server: see Add a node.
  • SSH access shows the saved SSH server and login, and Reveal password shows the password after a fresh sign-in confirmation; how the password is kept and changed is in Install a node over SSH.
  • Decoy site: what the node shows to anyone who is not a client; see Your own decoy site.
  • AmneziaWG backend: how the node runs AmneziaWG; see AmneziaWG.
  • Host: hostname, OS, kernel, architecture, CPUs, RAM, disk, virtualization, IPv6, boot time and the engine versions, as the agent reported them.
Field What it does
Name a–z, 0–9 and dash, 2–24 characters, unique.
Address The domain or IP clients connect to, without https:// or a port. Changing it changes the address in every subscription.
Country Two-letter country, or Not set. Server names in the apps use it ({flag} {country}); on nodes in Russia the doctor checks gosuslugi.ru and offers Yandex DNS, and on nodes in China it checks www.baidu.com and offers AliDNS and DNSPod.
Location Free text next to the name ("Frankfurt"). Empty shows the country name.
Public label Optional text, up to 40 characters, that people see next to this server in the Mistgate app ("Fast", "Home route"). The node's name is never shown to them; with no label they see the country and city alone.
Provider The hoster, for you.
Network capacity Mbps, optional: the provider's symmetric limit. The node's load is its CPU use, or the busier direction of its link as a share of this capacity when that is higher; without it the load is the CPU alone (on a one-core VPS the CPU usually runs out first). The server names in Happ end with the load (" · 64%"), and the user page shows a level: low, medium or high (from 50% and 80%). Never the rates themselves, and nothing in other apps, whose names must stay stable. 0 means unknown. Measure can fill it in: see below.
Notes Free text, up to 500 characters. The first two lines show in the Nodes list; user pages and subscriptions never show them.
Drop the agent connection after 30–600 s, default 90. If the agent stays silent longer, the panel closes the connection. The node turns grey (Host blip) first; Unreachable comes after 10 minutes without contact.
Wait for a command result 10–900 s, default 120: how long the panel waits before a command counts as unanswered.
Agent dial timeout 5–120 s, default 15: how long the agent waits when it connects to the panel.

Save sends only what changed, and the form on Exit saves its own fields, so neither overwrites the other. Most settings take effect without a reconnect.

Measuring the network capacity#

Providers often do not state the real channel, and a virtual server's network card does not report the plan's limit, so the node can measure how fast it reaches the internet. Next to Network capacity the owner presses Measure. One measurement is three runs back to back, a second apart, and the best run counts, per direction. In a run the agent on the node downloads for about 5 seconds over 6 parallel connections from a public speed server that needs no account, leaves out the first second (TCP is still speeding up), and then uploads for about 2 seconds over 4 connections to the same server. The agent picks the server in this order:

  1. The nearest Ookla (Speedtest.net) servers. The node asks the public Speedtest.net server list, which is geolocated by the node's own address, so it returns servers near the node (often inside its provider's network). The agent sends a short HTTPS request to each of them, and uses the one with the lowest latency, and the second-lowest if the first cannot serve. The test is GET /download?size=N and POST /upload on that server. These servers are nearly always closer, faster and less restricted than Cloudflare, which some countries throttle.
  2. speed.cloudflare.com, when the list or those servers do not answer. Cloudflare limits how often one address may ask (HTTP 429), so the runs on it are 3 seconds apart, a 429 is waited out (for Retry-After, at most 2 seconds) instead of ending the run, and every request asks for the most it allows, 50 MB.
  3. A file on proof.ovh.net, then on cachefly.net, when Cloudflare is not reachable either (download only).

The whole measurement takes about 30 seconds, moves at most 1.5 GB (per run up to 400 MB down and 100 MB up), and ends within 45 seconds whatever the servers do. Only one measurement runs on a node at a time.

The panel shows "Measured: 940 Mbps ↓ · 870 ↑" and, next to it, Use 870. It puts the slower direction, rounded (to 5 below 100, to 10 below 1000, to 50 above), into the field: a VPN node relays every byte both ways, and providers often leave the inbound free while they cap the outbound (a node can measure 5000 down and 1000 up on a 1 Gbit/s plan); without an upload figure, the download. The first measurement of a new node stores the same value; nothing is saved until you press Save, and the measurement never overwrites a value you typed by itself. Uploads are measured against Ookla and Cloudflare: if the node fell back to OVH or CacheFly, the upload is not measured.

The panel says how the runs went. Under the figures it names the server ("Through Ookla · MTS, Moscow", the sponsor and city of the Ookla server) and how many runs worked: "the best of 3 runs" when all did. When some did not, it says so and gives the reason: "1 of 3 runs worked: the others — the server limited the requests". The reasons are: the server limited the requests (HTTP 429), it did not answer in time, it did not answer at all, it answered with an error code, or there was no time left to start the run. Fewer successful runs make the figure less reliable, since the best of fewer tests is likely lower: press Measure again.

The people already on the node count. During the counted part of every run the agent also reads the byte counters of the node's main network interface (the same ones the node's load figures come from). The estimate of a direction in a run is the larger of what the test alone moved and what the whole interface carried in that window, so what the people using the node carry during the test is part of the capacity, not lost from it. When the interface carried more than the test did (allowing for the test's own protocol overhead), the panel adds "Including people's traffic: 35 Mbps ↓". That part is already inside the figures, not on top of them. If the counters cannot be read, only the test counts.

Take the number as an estimate, not as the provider's limit:

  • It is three short tests. It depends on the route from the node to that test server, which can be slower or faster than the route to your users.
  • It measures what the node can use now, and the users feel the test: their speed may dip for those 30 seconds.
  • A very fast link (above a few Gbit/s) hits the 400 MB limit of a run's download within a second or two and the result is a lower bound.
  • The result says which server answered and how many runs worked. If your provider states the limit, keep their figure.

Only the owner can measure (the API refuses tokens and MCP), and each measurement is written to the audit log (node.bandwidth_measure). It needs a node with the agent from this release or newer: for an older agent the panel says "agent too old" and sends nothing, and the node must be connected. The node needs outbound HTTPS to the test servers: www.speedtest.net (the list) and the Ookla servers it names (port 8080), or the fallbacks.

A new node is measured once automatically: about 45 seconds after its agent connects for the first time after the installation, if Network capacity is still 0. That is the only value the panel writes by itself, and only when the field is 0; the result is in the node's events ("Network capacity measured at the first start") and in the audit log (node.bandwidth_auto). The panel stores the slower of the two directions, and only if the test measured the upload as well: if only the download answered (the test servers refused the upload), nothing is written and the node gets a warning event ("The first measurement got ... no upload figure"), because on a link that is faster down than up the download alone would overstate the capacity. Press Measure and type the number yourself. A node whose agent is too old, or whose test fails, is skipped without a message: press Measure yourself or type the number.

Note

AmneziaVPN keys hold the node address they were issued with. After you change Address, the devices of the node's AmneziaWG profiles are marked outdated, and the people who connect with keys must import their key again (from their page or from the user card). Subscription apps get the new address at their next update. The IPv6 entry never touches these keys.

Renaming a node#

Change Name and save. The name shows in the panel, in the file names of AmneziaWG configs, and in app server names only when the template uses {node}. The default uses the country and profile name. See Subscriptions.

Danger zone: retiring a node#

Retire from fleet removes a node for good. The dialog lists what happens:

  • users stop getting the node in their subscription; their links stay the same;
  • the node's certificate is revoked and its agent is told to stop; the files on the server stay;
  • the node's record stays in the audit log, with its traffic history.

To confirm, type the node name. After that the node disappears from the list; its page, if you open it, says "This node is retired". A retired agent cannot connect again. Only the owner can retire a node.

A node installed over SSH keeps its SSH access card after retiring: the panel no longer changes that server, but you can still reveal the saved password, and Forget saved access (asked twice, with a step-up) deletes it for good. When the panel generated the saved password itself, the retire dialog warns you first, since only the panel knows it.

If the node was offline, its agent never got the order and keeps serving the users it had. The panel then stays on the node's page and says "{name} did not get the order", with the commands that clean the server by hand: the same as in "Remove a node" in Add a node.

There is no other way to delete a node: retiring is the removal. To clean the server afterwards, as root:

Shell
systemctl disable --now mistgate-node
rm /etc/systemd/system/mistgate-node.service /usr/local/bin/mistgate-node
rm -r /var/lib/mistgate-node

The paths are the defaults of mistgate-node install; if you installed with other --bin or --state-dir values, use those.

Agent version and drift#

  • Outdated agent. The panel knows the newest agent version it ships. A node that runs an older one is marked outdated agent in the list and in its header. Some features need a newer agent: the doctor, WARP, the automatic AmneziaWG module build. Update it from Updates.
  • Capabilities. An agent lists what it can do, and the panel sends a feature only to an agent that listed it. user-rate-limit/1 means the agent can share one speed limit among a person's devices on the node (Hysteria2 and VLESS). A node whose agent lacks it holds each device to the smaller of the person's and the device's limits, and the Agent card on the node's Server tab says so: "The agent on this node cannot share a speed limit between a person's devices…". Update the agent to get the per-person limit. See Users and groups.
  • State drift. The panel sends every node its full desired state and compares a hash of it with what the node applied. When they differ, the panel sends the state again (event "state drifted from the panel, resent"). If the node still applies a different state, it reads Needs attention with "the node's state differs from the panel's", and an alert opens. Restart the profiles; if it comes back, read the node's log.
  • Clock skew. When the server clock differs from the panel's, the node says so. A wrong clock breaks certificates and handshakes; turn on time sync on the server (timedatectl set-ntp true).

Who can do what#

Every admin sees the nodes. The owner and helpers can change a node's settings (except its address) and restart its profiles. Moving a node to another address, installing nodes over SSH, making install commands, revealing and forgetting saved SSH access, retiring nodes, reading logs, putting profiles on nodes, turning IPv6 exit off or on, checking and changing the IPv6 entry and building the AmneziaWG kernel module are for the owner only; so are relays and Add a way in… (see Ways in). See Security.

Edit this page on GitHub