AI assistant and MCP

The AI assistant and the MCP server are Pro edition features. In the Community edition these menus do not work and /api/mcp requests are refused. See Pricing for the edition differences.

#Two separate parts

Both use the same tool catalogue. Tools never act directly: every call goes to the panel's own API endpoints with the identity of the person using it. Permissions, package features, reseller scope and activity logging work exactly as in the rest of the panel, so the tools cannot be used to gain extra privileges.

#Settings

Settings live on Server Admin → System → AI & MCP (server administrator only). The page has three tabs: Assistant, Usage and limits and MCP.

#Provider and API key

The assistant works with two providers, using your own API key:

Keys are stored encrypted and never shown again; the interface only shows a short hint such as sk-ant-…ab12. The Test button checks the key against the provider. The selected provider applies to new conversations only: each conversation continues with the provider and model it started with. Removing the key turns the assistant off, while the MCP server keeps working.

#Assistant options

The Usage and limits tab shows today's usage per person and provider plus totals for the last 14 days, and lets you set a per-person daily limit by username (unlimited, or 0 to switch it off). Once the budget is used up the assistant makes no new requests that day. When an administrator is signed in to another account on the owner's behalf, usage counts against the administrator's budget; team members' usage counts against the account owner's budget.

#Using the assistant

Open the assistant as a drawer from the top bar or from the menu:

• User panel: Advanced → AI Assistant
• Server Admin: System → AI Assistant (server administrator and reseller)

For account owners it can be turned off per package with the AI assistant package feature. The assistant gathers facts with its tools to help with site problems, DNS and email issues, SSL status, backups and resource usage. It has no shell, file system or internet access; it can only use the listed tools.

Read-only tools run immediately. Tools that change something (create, update, delete, suspend, restart a service, issue a certificate, start a backup) show a confirmation card and run only after you approve; writing a new message while a card is waiting declines the pending calls. Every approved change is written to the activity log. One reply takes at most 12 tool steps; if it hits the limit, type "continue" to let it go on.

#Data sent to the provider

Chat messages and the results of the tools the assistant calls (domain settings, DNS records, log lines, account summaries…) are sent to the conversation's provider to generate replies. Before sending, passwords, private key blocks, API tokens, JWTs, cloud access keys and credentials in URLs are replaced with [redacted]. Conversations are stored in the panel database in the same redacted form for the retention period, and each person sees only their own conversations (team members and the account owner cannot see each other's).

#MCP server

The Model Context Protocol (MCP) server lets AI applications that support MCP (Claude Desktop, Claude Code, OpenAI Codex and others) connect to the panel. There are two transports:

• Streamable HTTP: the /api/mcp endpoint on your panel address, with Authorization: Bearer <API_TOKEN>.
• stdio: the bipanel mcp command on the server. It forwards messages to the /api/mcp endpoint of the running panel and uses the server's CLI token by default, which means full server administrator rights.

The server offers tools, summary resources (account summary, server status, account details) and two prompts (site troubleshooting, security audit). Open sessions are listed on the MCP tab; sessions idle for 24 hours are closed.

#MCP options

For account owners it can also be turned off per package with the MCP access package feature.

#Tokens and permissions

MCP tools run with the permissions of the token owner; tokens have no separate scope setting. Create a dedicated token for MCP and give it an expiry if you like (30 days, 90 days, 1 year or none). Tokens are created and revoked on Advanced → API Tokens (user panel) and System → API Tokens (Server Admin); a token is shown only once, when it is created.

Tools that make changes run only when the client repeats the call with "confirm": true. Make sure your client asks you to approve these calls. Secrets in tool results reach the MCP client as [redacted] as well.

#Connecting a client

The panel shows connection instructions with ready-made configuration: Advanced → MCP Connection in the user panel, System → MCP Connection for resellers, and the MCP tab of AI & MCP for the server administrator. The page shows the full endpoint address and two formats:

• Claude Desktop and Claude Code: the mcpServers JSON format; for HTTP, type: "http", the url and an Authorization header.
• OpenAI Codex: a [mcp_servers.<name>] section in ~/.codex/config.toml. The token is not written to the file; it is read from an environment variable through bearer_token_env_var, and the same entry can be added with codex mcp add.

The server administrator can use stdio on the server itself with bipanel mcp, or from another computer over SSH:

$ ssh root@server bipanel mcp

To run it with another account's permissions, add --token <API_TOKEN> to the command. If the panel rejects the CLI token, regenerate it with sudo bipanel admin reset-token. In a Docker or Railway container the bipanel command lives inside the container, so the HTTP endpoint is the most direct way to connect from outside (Docker and Railway).

#Related pages

• API and CLI
• Accounts and packages
• Resellers, teams and branding
• Panel security