Skip to content
EN
Русский English
Connect a server

Manage your server with words.

The assistant knows your infrastructure: it reads logs and metrics, deploys, runs commands and SQL. Changes wait for your approval, and every call lands in the audit log.

Connect a server All features
  • 7 days free
  • No card required
  • Payment in rubles

Assistant chat

Why is shop.example.com returning 502?

Checking logs and metrics…

tail_log server_metrics read_configuration

PHP-FPM hit its limit: pm.max_children = 16, all workers are busy. I suggest raising it to 32 and restarting the pool.

Approval needed Deny Approved

update_php_settingspm.max_children = 32sudo

restart_servicephp8.4-fpm

Done. Pool restarted, shop.example.com is returning 200 again.

MCP

Reads3

no approval

Changes2

Approved

Server web-01

shop.example.comweb-01 · 203.0.113.10

200OK

38 msResponse time

tail_log

Logs · nginx-error

[error] connect() to unix:/run/php/php8.4-fpm-shop.example.com.sock failed (11: Resource temporarily unavailable)

[error] connect() to unix:/run/php/php8.4-fpm-shop.example.com.sock failed (11: Resource temporarily unavailable)

no new entries
server_metrics

Metrics

CPU34%

RAM61%

Disk48%

read_configuration update_php_settings restart_service

PHP-FPM 8.4Workers 9 / 32

pm.max_children32

80
panel tools — the same ones the MCP server exposes
3
access levels: read only, operate and sudo
4
providers to choose from when you bring your own key
200
AI actions per month on Pro and 5,000 on Business

Tools

Ask — and it gets done.

The assistant doesn't just answer — it calls real panel tools and shows every step. Everything you do by hand in the panel, the assistant does with a tool call.

Deploy

Ships a commit, waits for the task to finish and re-runs a failed one.

Calls deploy_website wait_for_task restart_task

Logs

Reads the tail of a log and quotes the line with the error.

Calls list_logs tail_log

Metrics

Server CPU, memory and disk, service status and whether the site answers.

Calls server_metrics get_server website_status

Databases

Read-only SQL, dumps and writes — as a separate MySQL user with access to one database only.

Calls db_query db_dump db_execute

Commands

Shell, artisan and WP-CLI — on the server or in the site directory.

Calls run_command run_artisan run_wp_cli

Backups

Runs a backup now, lists earlier copies and restores the one you need.

Calls run_backup list_backups restore_backup

SSL

Issues a certificate for a new domain and renews an expiring one.

Calls install_ssl renew_ssl

Firewall

Shows UFW rules, opens or closes a port.

Calls list_firewall_rules create_firewall_rule

Control

It acts. You decide.

The assistant goes through the same checks as MCP tokens: access level, approvals, the dangerous command guard and secret masking.

  1. Approvals

    It changes the server only at your click

    The assistant reads on its own. Before a deploy, a restart or a database write it stops and shows a card: what will run and where. That is the default, and for sudo actions it is always the case.

    • Logs, metrics and SELECT queries run without stopping
    • Several actions in one card, and you can remove any of them
    • Command arguments are frozen: the model can't swap them after you approve
    Approval needed3 actions · shop.example.com
    Done
    deploy_websiteshop.example.com · main
    done · 14 s
    restart_servicephp8.4-fpm
    done · 2 s
    run_site_command sudonpm ci && npm run build
    done · 19 s
    Approved Deny
  2. Access level

    How far it can go is set in settings

    The assistant has three levels: read only, operate and sudo. Operate is the default: deploys and restarts are in, arbitrary shell and database writes are out.

    • Sudo tools are enabled one by one
    • Sudo can be enabled only after you confirm your identity — with a password or a security key
    • The assistant never gets more rights than you have in the team
    Access levelShell, SQL writes, firewall
    sudo enabled
    Read-onlyManageFull (sudo)

    Sudo tools

    run_commandArbitrary shell on the server
    db_executeWrites to a database
    create_firewall_ruleOpen or close a port
  3. Dangerous command guard

    A destructive command never reaches the server

    Every shell command and every SQL statement is checked before it is sent. The guard is on by default; a refusal goes to the audit log and you get a notification.

    • rm -rf, dd, shred, reboot, disabling the firewall
    • DROP, TRUNCATE, DELETE and UPDATE without WHERE
    • Deleting inside a site's own directory is allowed — old releases can be cleaned up
    • Disk formatting and rm -rf / are always blocked and cannot be switched off
    Dangerous command guardChecked before it reaches the server
    rm -rf releases/old-*allowed: inside the site directory
    rm -rf /var/wwwblocked: recursive rm
    DROP DATABASE shopblocked: DROP DATABASE
    DELETE FROM ordersblocked: DELETE without WHERE
    mkfs.ext4 /dev/vda1always blocked: disk formatting
    Blocked4
  4. Secrets hidden from AI

    Passwords and keys never reach the model

    Before an .env, a config or command output goes to the model, secret values are replaced with [redacted]. The same applies to external MCP clients.

    • By key name: PASSWORD, SECRET, TOKEN, KEY, DSN
    • By the shape of the value: a password in a URL, a JWT, a bot token, a generated key
    • Arguments with passwords and tokens are masked in the audit log too
    • SQL runs as a separate MySQL user — database passwords never leave the server
    read_envshop.example.com · .env
    3 masked
    Tool responseMask
    APP_ENV=production
    APP_URL=https://shop.example.com
    DB_HOST=127.0.0.1
    DB_PASSWORD=[redacted]
    PAYMENT_SECRET=[redacted]
    REDIS_URL=redis://:[redacted]@127.0.0.1:6379
    The file as the model sees it

Audit log

Every call is on the record.

The assistant never works around the panel: every tool call is recorded, allowed or denied. One log covers the chat, the Telegram bot and MCP clients.

  • Audit log: who, when, which tool, result
  • Filter by denied calls and by token
  • Changes to the assistant's access level are recorded too
  • Allowed calls are kept for 30 days on Pro and 365 on Business
MCP activityEvery tool call, including denied ones
ToolDecision
tail_logAI assistant · Anna · 3 min ago
allowed
server_metricsAI assistant · Anna · 3 min ago
allowed
restart_serviceAI assistant · Anna · 1 min ago
allowed
run_commandclaude-desktop · dangerous command guard
denied
deploy_websiteAI assistant · Anna · just now
allowed
38AI actions
1denied
30days kept

Telegram bot

The same assistant — right in Telegram.

Text the bot in plain words — and it deploys, reads logs, runs a command or SQL. The same assistant, same tools and same limits as in the panel. The reply arrives in real time.

  • In the panela chat next to your server
  • Telegramfix a site from your phone
  • Claude · Cursorand any MCP client
  • Approve dangerous actions with a button right in the chat — marked SUDO
  • Alerts land here too: expiring SSL, a failed deploy, an unreachable server, a full disk
  • Connect with a single link from the panel settings

MCP server for Claude and Cursor

Deploykinbot
Deploy main to prod and clean up old releases

Deploy main → prod started…

deploy_website wait_for_task

Deploy finished in 14 s. I will remove old releases with a command — it needs your approval.

Approval needed · SUDOrm -rf releases/old-*
Approved

Done. Removed 6 old releases, freed 1.2 GB.

replies in real timeapproval with a buttonalerts

Model and limits

Works out of the box. Or on your own key.

On the Pro plan the assistant runs on Deploykin AI — no keys, no setup. On Business, connect your own key from your favorite provider and pick the model: tokens are billed directly to them, with no markup from us. Conversation history is saved in the panel.

Your own key

Your provider, your model

On Business, pick a provider, paste a key and test it with one button. You can switch the provider or the model at any time.

  • Anthropic, OpenAI, Google Gemini and DeepSeek
  • The key is tested before it is saved
  • Model tokens are billed by the provider directly
ProviderSettings → AI assistant
Key works
Anthropic OpenAI Gemini DeepSeek

Key and model

API key
sk-••••••••91c7
Model
gpt-5.1
Test Save
Limits

Actions are counted, not questions

An AI action is a tool call that changes something: a deploy, a restart, a database write. Reading logs and metrics and running SELECT queries do not use up the limit.

  • 5,000 actions per month on Business, 200 on Pro, 50 during the trial
  • One limit for the assistant, the Telegram bot and MCP clients
  • When the limit runs out, reading and diagnosing still work
AI actionsPro · resets on the 1st
162 left

Used38 / 200

CallCharged
tail_log
log read
0
db_query
SELECT query
0
deploy_website
site deploy
+1
On the Business plan5,000 per month

In the panel 203.0.113.21 Connect done in 4 min 52 s

Hand the AI the keys to your routine.

The assistant and MCP server are available on the Pro and Business plans. Open the chat in the panel — and start managing your infrastructure with words.

Start for free View pricing
  • 7 days free
  • No card required
  • Human support

We use cookies and the Yandex Metrica web-analytics service (including Session Replay) to improve the site. Details are in the privacy policy.