Agents & API¶

Mailr has one API. The web app, the Android app, the built-in assistant and external agents are all built on it, so every action available in Mailr is also available to an agent, under the same rules.

An agent can be connected in three ways:

Way in	Best for	Page
MCP server	Claude Code, desktop assistants and any MCP client	MCP for Claude Code and other agents
Tool API	scripts that call one tool at a time over HTTP	Agent tools
REST API	apps and integrations built on the full API	REST API

All three require an API token created in Mailr. See Tokens and scopes.

Settings, Agents & API, with tokens, permissions and the approval queue

Access and safeguards for agents¶

The same tools as the built-in assistant: search by words or by meaning, read threads, people profiles, a context bundle for replies, drafts, labels, filters, calendar, tasks, memory, skills and Speakr recordings.
Scoped access. Access through a token is limited to the accounts and actions granted to it.
Local-only protection. Mail from a local-only account is invisible to a token that is not marked local-only.
A send gate. Agents write drafts by default. Sending, deleting and changing rules are held for approval unless a rule allows them. The first message to a new address always requires approval. See Approvals for agents.
An audit log of every call, under Settings → Agents & API → Audit log.

Mail is untrusted input¶

Email content is marked as untrusted in every response. Agents must treat it as data, never as instructions. The rule is stated in the tool descriptions and in the instructions given to each agent by the MCP server.