Skip to content

Your infrastructure, open to AI agents.

Claude, Cursor or your own agent get 80 panel tools without a single SSH key. You decide what a token may do: which servers, what access level and for how long. Every call is audited.

  • 7 days free
  • No card required
  • Payment in rubles
All 80 tools of the MCP server
Claude Desktop

Deploy main to prod and clear the config cache

Token
  • add_repository
  • adopt_server
  • attach_server_to_project
  • cancel_task
  • check_system_updates
  • configure_system_updates
  • control_cron_job
  • control_daemon
  • copy_backup_provider
  • create_backup
  • create_backup_provider
  • create_cron_job
  • create_daemon
  • create_database
  • create_docker_container
  • create_firewall_rule
  • create_project
  • create_server
  • create_website
  • create_worker_node
  • db_dump
  • db_execute
  • db_query
  • delete_backup
  • delete_database
  • delete_website
  • deploy_from_archive
  • deploy_website
  • generate_server_key
  • get_deploy_script
  • get_server
  • get_task_output
  • install_package
  • install_ssl
  • list_backup_providers
  • list_backups
  • list_configurations
  • list_cron_jobs
  • list_daemons
  • list_firewall_rules
  • list_logs
  • list_projects
  • list_repositories
  • list_servers
  • list_tasks
  • list_teams
  • migrate_from_ploi
  • read_configuration
  • read_env
  • read_file
  • register_log
  • register_service
  • renew_ssl
  • restart_service
  • restart_task
  • restore_backup
  • resume_website_provisioning
  • run_artisan
  • run_backup
  • run_command
  • run_site_command
  • run_system_updates
  • run_wp_cli
  • server_conventions
  • server_metrics
  • set_deploy_sync
  • sync_repositories
  • tail_log
  • test_server_connection
  • transfer_server
  • update_backup
  • update_deploy_script
  • update_mysql_settings
  • update_nginx_settings
  • update_php_settings
  • upload_file
  • wait_for_task
  • website_status
  • write_configuration
  • write_env
  • +58

website_statusHTTP 200 · nginx -t ok

read operate sudo databases have their own level

80
panel tools: from deploys and logs to SQL and the firewall
3
access levels — read, operate and sudo — for every server and site
0
SSH keys and passwords held by the agent: it only has an MCP token
4
clients with a ready-made guide: claude.ai, Claude Desktop, Claude Code and Cursor

Use cases

What teams automate over MCP.

80 tools: everything you do in the panel by hand is available to an agent. On the right — the chain of calls and the access level the token needs.

Deploy from the editor

“Deploy this branch and wait for the result” — right from Cursor, without switching to the panel.

  1. deploy_website
  2. wait_for_task
  3. restart_task
operate

Incident triage

“Why 502?” — the assistant reads the logs and metrics itself and suggests a fix.

  1. tail_log
  2. server_metrics
  3. website_status
operate

Database work

“Dump the database before the migration and show slow queries” — SQL without a console.

  1. db_dump
  2. db_query
select-only

Commands without SSH

artisan, WP-CLI and arbitrary commands — as the site's user, with an audit trail.

  1. run_artisan
  2. run_wp_cli
  3. run_site_command
operate · sudo

Server and site from scratch

“Connect this VPS, install the Laravel stack and create a site from the repository” — the agent starts provisioning and waits for it to finish.

  1. create_server
  2. adopt_server
  3. create_website
  4. install_ssl
sudo

Ready-made runbooks

Besides tools, the server ships two checklists — a safe deploy and a server diagnosis — plus a usage guide that tells the agent how to work with the panel.

  1. safe_deploy
  2. diagnose_server
  3. usage-guide
prompts

Control

Freedom for the agent, control stays with you.

Between “the AI decided” and “it ran on the server” stand the token's level, its owner's role, the list of allowed tools and the dangerous command guard.

A token instead of keys

The token decides where an agent may go and how deep

The agent needs no keys, passwords or server credentials. It gets an MCP token — and the panel performs every operation itself over its own secured connection. Revoke the token and access closes instantly, with no keys left on servers to clean up.

  • Servers and sites get read, operate or sudo; databases get read, select-only or full
  • A token can never do more than the person who issued it — the team role stays the ceiling.
  • A sudo token is issued with password confirmation and always expires — in an hour to a year
  • The secret is shown once: the panel stores only its hash
New access tokenclaude-desktop
Created
Access scopesLevel
prod-01server · its sites and databases
operate
blog.example.comsite on staging-01
read
analyticsdatabase on db-01
select-only

Expiry

7 days 30 days No expiry
Cancel Create token
Copy it now — shown only oncedk_mcp_x7Lq2Rv…
Sudo behind a key

Sudo opens with a touch of your key — and locks itself

A token's read and operate always work. Its sudo can be locked behind a security key: you open it from the token list for 15 minutes, an hour or eight, and it locks again when the time is up.

  • A locked sudo answers the agent with a denial and a hint to ask the owner to unlock it
  • Lock sudo early with a single button
  • The token list shows whether sudo is locked or until when it is open
  • An OAuth connection from claude.ai never gets sudo: it is capped at operate
Access tokensSettings → MCP
14:32
TokenLevel
claude-desktopsudo locked
sudo
cursorall servers · no expiry
read
OAuth · Annaclaude.ai · by role
operate

Open sudo for

15 minutes 1 hour 8 hours
Open Lock sudo
Locked itself after 15 minutes
Only what's needed

Each token has its own list of tools and addresses

The level sets how deep a token can go; the tool list sets what exactly it can do. Disabled tools disappear from tools/list, so the agent never learns they exist.

  • Four core tools are always on — without them an agent cannot find a server or read a task's result
  • IP restriction: from any other address the token cannot even list the tools
  • A call to a tool outside the list is recorded in the audit log as denied
Restrict toolsclaude-desktop · operate
list_serverscore — always on
deploy_websiteRun a website deploy
tail_logRead the tail of a log
db_queryRead-only SQL
restart_serviceRestart a service

Where the token works from

203.0.113.10
others are refused
The agent sees in tools/list7 / 80
Audit and alerts

An agent is denied — the team knows right away

The activity log records every call — allowed or denied — with the token name and the reason for the denial. Denied calls trigger an alert to your subscribed channels.

  • At most one alert per token per 15 minutes — a looping agent will not flood the chat
  • Allowed calls are kept for 30 days on Pro and 365 on Business, denied ones for 180 days
  • Passwords and tokens in call arguments are redacted, and the agent gets .env and config files with secrets masked
  • The dangerous command guard blocks rm -rf, mkfs and DROP DATABASE before anything is sent to the server
Denied callsAlerts go to subscribed channels
Call · claude-desktopTime
run_commandneeds sudo, the token grants operate
14:07
db_executeneeds full, the token grants select-only
14:09

Channels

Telegram
sent · 14:07
Email
sent · 14:07
Next alert for this token — not before14:22
1 / 4
Pricing

MCP — on the Pro and Business plans

Reads — metrics, logs, statuses — don't consume the limit: only mutating actions count.

  • The limit is shared by the assistant, the Telegram bot and MCP clients
  • Over the limit, you can still read and diagnose
  • The counter resets at the start of each month
MCP serverLimits by plan
LimitProBusiness
MCP tokensactive
2 Unlimited
AI actionsper month
200 5 000
Auditretention
30 days 365 days

Free of charge

list_servers get_server server_metrics tail_log db_query read_env website_status wait_for_task
An AI action is a call that changes something+1

Quickstart

Three steps to your first “deploy it”.

MCP is an open standard: one endpoint for everyone. The panel has step-by-step guides for claude.ai, Claude Desktop, Claude Code and Cursor — and a prompt that lets any other agent connect itself.

  • Issue a token in the panel and set scopes: read, operate or sudo — for a whole server, a single site or a database.
  • Add the endpoint to your client: copy the snippet from the panel with your token already in it.
  • “Deploy staging”, “why is prod returning 502?”, “dump the database” — the client calls the right tools itself.
  • Using claude.ai? Connect Deploykin via OAuth — no token copy-pasting.
Connect a clienthttps://app.deploykin.com/mcp
Copy
claude.ai Desktop Code Cursor
Where it goes~/.cursor/mcp.json
# claude.ai → Settings → Connectors
# → Add custom connector
https://app.deploykin.com/mcp

# no token needed: sign in with OAuth
# level follows your role, up to operate
"mcpServers": {
  "deploykin": {
    "url":
      "https://app.deploykin.com/mcp",
    "headers": {
      "Authorization":        "Bearer dk_mcp_x7Lq2Rv…"    }
  }
}
$ claude mcp add --transport http \
    deploykin \
    https://app.deploykin.com/mcp \
    --header \    "Authorization: Bearer dk_mcp_…"
"mcpServers": {
  "deploykin": {
    "url":
      "https://app.deploykin.com/mcp",
    "headers": {
      "Authorization":        "Bearer dk_mcp_x7Lq2Rv…"    }
  }
}
Your fresh token is already in the snippet

Task manager

Every AI action is a task you can see.

The agent does nothing “around” the panel: every action becomes a task in Deploykin's shared task manager — right next to your own deploys and operations.

  • Every task has a status, output and author — tagged “MCP” or “AI”
  • The agent waits for results itself: wait_for_task, get_task_output
  • Emergency stop: cancel_task kills the process right on the server
  • A failed chain can be restarted from the first failed step
Tasksprod-01 · people, the assistant and MCP clients in one list
TaskStatus
Deploy shop.example.comMCP · claude-desktop
Finished · 41 s
Dump of the shop databaseMCP · cursor
Finished
Restart php8.4-fpmAI · Anna
Finished
Package updateAnna
Finished
wait_for_taskfinished

FAQ

FAQ about MCP.

The protocol, the safety model and how the MCP server differs from the assistant in the panel — in short.

What is MCP?

Model Context Protocol is an open standard that AI clients use to connect to external services. Deploykin runs an MCP server: any compatible client gets 80 panel tools — from deploys to SQL.

Does the agent really need no SSH keys?

Really. The agent only sees an MCP token. The panel performs all operations over its own secured server connection, SQL runs under separate managed database users, and .env files and configs are served with secrets redacted.

What if the AI asks for rm -rf /?

The command never reaches the server: the dangerous command guard blocks destructive patterns — rm -rf /, mkfs, fork bombs, DROP DATABASE, UPDATE without WHERE — before execution. The critical protection tier cannot be turned off. And arbitrary commands are only available to a sudo-level token, which is issued with password confirmation and always has an expiry.

Is there a classic REST API?

Today programmatic access to Deploykin is MCP: an open protocol over HTTP that works without AI too. If you specifically need classic REST — write to us and we'll share our plans.

How is this different from the built-in AI assistant?

The tools, access levels, guard and audit are the same. The assistant works right in the panel and in Telegram and shows an approval card before it changes anything. The MCP server exposes the same tools to external clients — Claude, Cursor and your own agents — where the token sets the boundaries.

In the panel 203.0.113.21 Connect done in 4 min 52 s

Give AI a token, not an SSH key.

7 days free. An MCP token takes a minute to issue in the panel.

  • 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.