There are two ways to put a reasoning model in front of your server: the built-in chat, and the MCP door for external agents. Both read and act through the same boundary — grants, approvals, and audit apply no matter which surface asked.
#The built-in chat
Chat targets your server and answers from the warehouse: mail, calendar, contacts, CRM snapshots, files, timelines. Point it at a model in Settings → AI/Models — either a free-pool key or a provider you bring.
When chat wants to act — send, update, create — the action arrives as an approval with editable arguments. Approve exactly one action, approve the batch it belongs to, or open a session-scoped grant; see Grants and approvals.
#Connect an external agent
Claude Desktop, Cursor, or any MCP-capable agent can connect to your server and use your tools. The MCP door manages its own credential — you never hand-author a token:
- Open the MCP door and choose the tools it exposes.
- Configure the agent with your server's MCP endpoint and the door's credential.
- The agent's calls arrive attributed to that agent, pass the same grants, and land in the same audit log.
What an agent can reach is exactly what the door grants: warehouse reads, timelines, and installed recipes as tools. Closing the door revokes its credential. The agent never sees your account credentials — it asks Recued to act, and Recued acts on your accounts under your grants.
MCP is not the only way in: chat apps that take a custom base URL and API key connect through the server's OpenAI-compatible endpoint. The client-by-client detail is in Connect to Recued.
#Chat platforms
Slack and Telegram connect as messenger doors. Messages and slash commands POST directly to your server — never through Recued's cloud — and replies, notifications, and approvals can reach you in the same channel.
#The trust model in one paragraph
An external agent is just another actor. It cannot see credentials, cannot bypass the boundary, and cannot approve its own asks — approvals go to you. Every call it makes is attributed and audited, and a running operation can be killed from your side at any time.
The same door machinery also works in the other direction: you can sell or grant tiered access to your own tools and AI to outside customers — see Selling access.