llamaindex-multimail

LlamaIndex tools for MultiMail -- give your LlamaIndex agents email capabilities with graduated human oversight.

Installation

pip install llamaindex-multimail

Quick Start

from llama_index.core.agent import ReActAgent
from llama_index.llms.openai import OpenAI
from llamaindex_multimail import MultiMailToolSpec

# Create tools from your API key
tool_spec = MultiMailToolSpec(api_key="mm_live_your_api_key")
tools = tool_spec.to_tool_list()

# Create agent
llm = OpenAI(model="gpt-4o")
agent = ReActAgent.from_tools(tools, llm=llm, verbose=True)

# Run
response = agent.chat("Check mailbox 'mbx_123' for new emails")
print(response)

Available Tools

Tool	Description
send_email	Send an email (held for approval if gated)
check_inbox	List recent emails in a mailbox
read_email	Read the full content of an email
reply_email	Reply to an existing email thread
search_contacts	Search contacts by name or email
list_pending	List emails awaiting human approval
decide_email	Approve or reject a pending email
get_thread	Get all emails in a conversation
tag_email	Add tags to an email for classification

This toolkit complements MultiMail's 38 MCP tools with LlamaIndex-native wrappers for common email workflows.

Compliance

MultiMail handles regulatory compliance at the infrastructure layer — no SDK-side code changes needed:

• EU AI Act Article 50: Every AI-sent email includes a cryptographically signed ai_generated disclosure in the X-MultiMail-Identity header
• US State Laws: Maine, New York, California, Illinois — AI disclosure built into email delivery
• CAN-SPAM: Unsubscribe headers and physical address footers on all outbound email
• Formally Verified: Lean 4 proofs of identity header tamper evidence

MultiMail handles EU AI Act Article 50 compliance at the infrastructure layer. Every AI-sent email includes signed ai_generated disclosure automatically.

See multimail.dev/use-cases/eu-ai-act-email-compliance for details.

Oversight Modes

MultiMail supports graduated oversight so your agent doesn't send unsupervised email:

• gated_all -- Agent drafts, human approves everything
• gated_send -- Agent reads freely, human approves outbound (default)
• monitored -- Agent sends, human can review after
• autonomous -- Full agent control

When a mailbox uses gated oversight, send_email returns pending_send_approval and the email waits for human review. The agent can check status with list_pending.

Using with Other LlamaIndex Patterns

The MultiMailToolSpec follows the standard LlamaIndex BaseToolSpec pattern, so it works with any agent framework that accepts LlamaIndex tools:

from llamaindex_multimail import MultiMailToolSpec

spec = MultiMailToolSpec(api_key="mm_live_your_api_key")

# Get all tools
tools = spec.to_tool_list()

# Or get a specific tool
send_tool = spec.to_tool_list(func_to_metadata_mapping={
    "send_email": {"name": "send_email", "description": "Send an email"},
})

Links

• MultiMail -- Homepage & docs
• multimail -- Base Python SDK
• langchain-multimail -- LangChain integration
• MCP Server -- For Claude, Cursor, and other MCP clients