Everything your agents can reach.
35 native connectors built into Cubebot, any MCP server you add, and six AI providers. You connect an account once; you decide which agents may use it, and how far they may go.
Connector catalog
35 connectors
- Painted tested against the live service
- Half-painted tested against recorded vendor responses
Communication
Calendar and meetings
Docs and files
- GoogleGmail, Calendar, Drive, Contacts, Sheets and Docs through your own OAuth client.
- Microsoft 365Outlook mail and calendar, OneDrive and To Do through your own Entra app.
- NotionSearch, read and write pages and databases.
- Nextcloud FilesBrowse, search, download, upload and organize files.
- Paperless-ngxSearch, tag, download and file your documents.
- DropboxBrowse, search, download, upload and move files.
Developer
Projects and CRM
Home and feeds
Notifications
Data
An MCP server or the HTTP API connector may cover it. Or ask for it on GitHub.
You set how far each account goes.
Every connection has a permission level, and agents are only offered the tools at or below it. Switch single tools off, share a connection with every agent or only the ones you choose, and connect several accounts of the same kind, like work and personal mail.
- Your own apps. Sign-in uses your API key or your own OAuth app. Cubebot registers no app of its own between you and the vendor.
- Short-lived credentials. Agents fetch a credential for each call. Refresh tokens and client secrets never leave the orchestrator, and secrets are struck from every result.
- Events start routines. New mail, an assigned issue or a finished recording can wake an agent. Events are polled, not pushed.
- Untrusted by default. Mail, documents and messages reach the model marked as outside content that cannot grant tools.
- Read
Look things up: search mail, read issues, list files. Nothing changes.
- Read and write
Also change things: create issues, file documents, add events.
- Full
Also reach people or move money: send mail, post messages, refund.
And anything that speaks MCP.
Add a Model Context Protocol server once under Platform › MCP servers. Local servers run inside each agent's sandbox as its tool user; remote servers are reached over Streamable HTTP. Secrets go in as encrypted environment variables or headers, and only their names are ever shown.
npx -y @modelcontextprotocol/server-githubFive AI providers, two plans.
Agents call the model themselves, with a key or plan token the orchestrator hands them for a short time. One provider and model run the whole installation at a time; switch between saved keys whenever you like.
- OpenAIResponses API, native computer use
- AnthropicPrompt caching, native computer use
- Google GeminiContext windows read from the API
- xAI GrokChat completions
- OpenRouterMany vendors behind one key
- MiniMaxAnthropic-compatible API, global or mainland China
- ChatGPT Plus and ProSign in with ChatGPT, a preview OpenAI offers to open-source apps
- SuperGrok and X PremiumOpt-in device sign-in; it identifies as the Grok CLI, so it carries an account-risk warning
- MiniMax M PlanThe plan’s own key, kept personal as MiniMax asks; pay-as-you-go is for shared use