Skip to content

Assistant

github-actions[bot] edited this page Nov 7, 2026 · 8 revisions

A cloud model in the admin that answers about your nodes, people and errors and prepares changes you apply. Providers, where the data goes, memory, limits and safety.

Assistant

The assistant is a chat in the admin with a cloud model you connect with your own API key. Ask it which nodes had problems, why someone cannot connect, who is over the device limit; ask it to make a group, change the subscription texts or turn a profile off for a node. It looks things up with the panel's own tools and prepares every change as a card in the panel's words. Nothing changes until you press Apply.

It works with the rights of the admin who talks to it, never more, and it is available on the VPS edition (see In the Cloudflare edition).

Turn it on

The owner sets it up in AI → Assistant:

  1. Under Providers, choose Connect on a provider, paste its API key and choose Check and save. The panel asks for your passkey or code, checks the key with the provider and seals it; the card then shows only its last four characters.
  2. On the connected card, pick the Model (the price per million tokens is under it) and the Monthly budget, $.
  3. Above, turn The assistant is on, pick the Provider the chat uses, How hard it thinks (Quick, Normal, Thorough), Who may use it (Only me, Owner and helpers, Everyone in the panel), the Notes for the assistant and how long to Keep conversations, and choose Save.

A round button appears at the bottom corner of every page for the admins it is open to; Ctrl+J (⌘J on a Mac) opens and closes the chat too.

Providers and where your data goes

Provider Address the panel calls Models Where the data goes
Anthropic api.anthropic.com from the provider's list; Claude Opus 5.5 by default, Sonnet 5.5 and Haiku 5.5 cheaper Processed in the United States; API data is not used for training.
DeepSeek api.deepseek.com from the provider's list Stored in China; the provider's policy allows using what you send for training.
Z.ai api.z.ai (the international platform) GLM-5.3 and GLM-5.3-Flash Processed in Singapore; the provider's agreement says content is not stored.
Ollama Cloud ollama.com the cloud models that take tools Processed in the United States; not used for training.

The card of each provider says the same, so you choose knowing it. The panel talks to a provider only from its own server, at that fixed address: there is no field for another address, and the browser never sees the key.

What the provider receives: the conversation, the assistant's instructions, your notes and the owner's notes, and the results of the tools it calls. Those results pass the same filter as MCP: no subscription links or page passwords, no keys, device configurations or tokens, no server passwords, no node addresses. If a message you type looks like a key or a password, the chat asks before sending it.

What it can do

  • Read: nodes and their health, people, groups, devices, events and alerts, reachability from the country, subscriptions and DNS, updates; the audit log for the owner. These are the MCP read tools of your role.
  • Change: every MCP change, as a plan. The plan comes to the chat as a Proposed change card with the panel's own summary, the facts (anything that came from data is quoted) and the danger notes. Apply runs exactly that plan, once; Dismiss drops it; a plan not applied within 10 minutes expires.
    • A plan that needs the owner (it changes what runs on the nodes, touches many people, or every person would see it): when the owner talks to the assistant, Approve and apply asks for the passkey or code and does both; for anyone else the card says it waits for the owner in AI → Waiting for you.
    • Installing a node over SSH is approved only in AI → Waiting for you, where the owner compares the host key and chooses how the panel logs in.
  • Its rights are yours: the owner's assistant has the Admin tools, a helper's the Operator tools, a read-only admin's only the read tools.

The model never applies anything itself: it has no tool to do so. Every call is in the audit log as "Assistant of ", with the source Assistant (filter by it in Settings → Audit).

Work out what does not open

Ask "who cannot connect, and what do we fix?", or press Work out what does not open on Health → People: the chat opens with the question ready to send. The assistant starts with access_report, which sets protocol against node (and group) over the last day: who tried, who got through, the rejects, stalls, freezes and loss the counters can tell apart, and who was given a server but never connected. It adds the view from the country (operators that see a node dark, which TSPU ports are cut, what kl!ck users saw) and block_diagnosis, which says how a node is blocked: the address is banned, one port is cut, TCP is throttled, UDP is cut, one protocol is recognised, only whitelisted addresses pass on mobile, or nothing is blocked and the node is simply broken.

It answers with a short picture per protocol, node and group, then three to five ranked recommendations, each tied to its evidence and offered as one plan (a relay through an entry people reach, a move to port 443, a different protocol for a group, a CDN profile, the IPv6 entry, taking a blocked server out of subscriptions). The rules are fixed, so the same facts give the same answer; where a signal is missing it says so and names the check that would tell (for example path_probe over UDP from a node in the country). VLESS keeps no attempt counter, so its failures show only through the country view and kl!ck.

Memory

  • Notes. The assistant saves what it should remember ("Remembered: …" in the chat, with Forget). AI → Memory lists them: add, edit and delete any, or Forget everything. Every note goes into every conversation; at most 100 notes of 400 characters.
  • Notes for the assistant (the owner's, in AI → Assistant): what every admin's assistant should know about this panel.
  • Conversations. The assistant knows the titles of your recent conversations and can search and read the earlier ones when you refer to them. AI → Memory → Conversations opens one in the chat, or deletes one or all.

Conversations and notes are sealed in the panel's database and private to each admin: nobody else reads them, the owner included. Conversations are kept as long as Keep conversations says; notes stay until you delete them. Whatever looks like a key or a password is cut out of everything the panel sends to the provider, apart from the message you type yourself.

Limits and cost

  • One answer uses at most 24 tool calls and 10 minutes; Stop ends it at once and keeps what was said.
  • One admin sends at most 20 messages a minute, and a conversation runs one answer at a time.
  • Each provider has the owner's Monthly budget, $, estimated from the provider's prices and the usage each answer reports. At the budget the assistant stops until next month. AI → Assistant → This month shows the cost per admin.

Safety

  • The key is sealed with the panel's master key, entered by the owner after a passkey or code, and never shown, logged or written to the audit log.
  • Names, notes and log lines from nodes and people are data to the assistant, never instructions; it has no tool that reaches the internet; the chat shows no images and no HTML, links to the admin's pages stay in the admin and other links open apart.
  • Even if a node name or a note talked the model into proposing something, it would only be a card: the change happens when a person reads it and presses Apply.

MCP, tokens and approvals

Outside agents (Claude Code, Cursor, your scripts) keep using MCP and API tokens: they moved from Integrations to AI → MCP and tokens, and the plans waiting for the owner to AI → Waiting for you, with the badge in the menu. Integrations keeps the Telegram alerts.

In the Cloudflare edition

The assistant is not available there yet, like MCP: AI → Assistant says so. It will come together with MCP.

When it does not answer

The chat says What to do
The provider refused the key. Replace the key in AI → Assistant.
The provider account has no balance. Top up the account at the provider.
This month's budget for the provider is spent. Raise Monthly budget, $, or wait for next month.
The owner has not opened the assistant to your role. The owner changes Who may use it.
Too many requests right now. Wait a minute; the provider or the panel limits the rate.
The provider did not answer, or is overloaded. Try again later, or switch the provider in AI → Assistant.

Edit this page · View on mistgate.app

Mistgate

English

Getting started

Guide

Operations

Roadmap

Reference

Русский

Начало работы

Руководство

Эксплуатация

План развития

Справочник

简体中文

入门

使用指南

运维

路线图

参考

Clone this wiki locally