MCP for agents

SentFromAI ships an open Model Context Protocol (MCP) server — sentfromai-mcp on npm — so an AI agent can use email as first-class tools: create inboxes, send and reply, search, manage threads, webhooks and allow/block rules — without writing a single API call. The agent simply calls a tool.

Why MCP

With the REST API your code orchestrates email. With MCP, the agent itself decides when to check its inbox, draft a reply, or escalate — SentFromAI’s operations appear directly in the model’s tool list.

Connect

All you need is an API key (sf_live_…) from the dashboard — the same keys as the REST API (Authentication).

Claude Code
Claude Desktop / Cursor / JSON config
Gemini CLI
Codex CLI
OpenClaw

As a plugin — asks for the key when you enable it and adds a skill that teaches Claude email etiquette:

claude plugin marketplace add sentfromai/sentfromai-mcp
claude plugin install sentfromai@sentfromai

Or as a plain MCP server:

claude mcp add sentfromai --env SENTFROMAI_API_KEY=sf_live_… -- npx -y sentfromai-mcp

Works for Claude Desktop, Cursor, Windsurf, Cline and any host that takes a JSON MCP config:

{
  "mcpServers": {
    "sentfromai": {
      "command": "npx",
      "args": ["-y", "sentfromai-mcp"],
      "env": { "SENTFROMAI_API_KEY": "sf_live_…" }
    }
  }
}

VS Code:

code --add-mcp '{"name":"sentfromai","command":"npx","args":["-y","sentfromai-mcp"],"env":{"SENTFROMAI_API_KEY":"sf_live_…"}}'

gemini extensions install https://github.com/sentfromai/sentfromai-mcp

Gemini asks for the API key on install and keeps it in your keychain.

codex mcp add sentfromai --env SENTFROMAI_API_KEY=sf_live_… -- npx -y sentfromai-mcp

openclaw mcp add sentfromai --command npx --arg -y --arg sentfromai-mcp --env SENTFROMAI_API_KEY=sf_live_…

Full walkthrough, including the OpenClaw skill: OpenClaw guide.

The server is a thin stdio wrapper over the REST API — no state of its own, and it runs wherever your agent runs (Node 20+). Source and issues: github.com/sentfromai/sentfromai-mcp; it is also listed in the official MCP registry as ai.sentfrom/mcp.

Hosted endpoint (remote MCP)

If your host only speaks remote MCP, or you would rather not run anything locally, connect to the hosted endpoint instead of npx:

https://api.sentfrom.ai/mcp

Same 21 tools, same tenant scoping, served over Streamable HTTP. Two ways to authenticate:

API key: send Authorization: Bearer sf_live_…. Works everywhere a bearer header can be set.
Sign in: hosts that support MCP OAuth (claude.ai, ChatGPT) discover SentFromAI’s authorization server automatically and open a sign-in to your SentFromAI account. The connection then acts for your organization’s tenant.

Claude Code (remote)
claude.ai / ChatGPT
Cursor / JSON config
Meta Muse

claude mcp add --transport http sentfromai https://api.sentfrom.ai/mcp --header "Authorization: Bearer sf_live_…"

Add a custom connector with the URL https://api.sentfrom.ai/mcp and complete the sign-in when prompted. No key to paste.

{
  "mcpServers": {
    "sentfromai": {
      "url": "https://api.sentfrom.ai/mcp",
      "headers": { "Authorization": "Bearer sf_live_…" }
    }
  }
}

Muse runs in Meta’s cloud, so use the hosted endpoint: add a custom connector pointing at https://api.sentfrom.ai/mcp with your API key as the bearer token.

The endpoint is stateless (one JSON-RPC request per HTTP request, no sessions). Tool calls run against the same API as the REST routes, so suppression, allow/block rules, plan limits and usage metering apply identically.

Available tools

Area	Tools
Inboxes	`create_inbox` · `list_inboxes` · `delete_inbox`
Messages	`send_message` · `reply_to_message` · `forward_message` · `get_message` · `search_messages`
Threads	`list_threads` · `get_thread`
Drafts	`create_draft` · `send_draft` · `list_drafts`
Domains	`add_domain` · `get_domain`
Webhooks	`create_webhook` · `list_webhooks` · `delete_webhook`
Allow/block lists	`add_address_rule` · `list_address_rules` · `delete_address_rule`

Every tool carries MCP annotations: reads are marked read-only, send_message, reply_to_message, forward_message and send_draft are marked as reaching the outside world, and delete_* as destructive, so hosts can group them and confirm the irreversible ones. API failures come back as tool results with the HTTP status (a 401 means the key is wrong, a 402 that the plan limit is reached), so the agent can tell you rather than retry. Because the tools map onto the same endpoints documented in the API reference, everything you read here applies — threading, idempotency, suppression, and attachments all behave identically.

Configuration

Env var	Required	Default
`SENTFROMAI_API_KEY`	yes	—
`SENTFROMAI_BASE_URL`	no	`https://api.sentfrom.ai`

Tools inherit the key’s tenant scope: an agent can only see and act on its own tenant’s inboxes, threads, and rules. Suppression and allow/block rules are enforced server-side, so a misbehaving agent can’t mail past them.

Get started

Core concepts

Guides

Examples

Agent frameworks

DNS guides

Troubleshooting

Why MCP

Connect

Hosted endpoint (remote MCP)

Available tools

Configuration

​Why MCP

​Connect

​Hosted endpoint (remote MCP)

​Available tools

​Configuration

Why MCP

Connect

Hosted endpoint (remote MCP)

Available tools

Configuration