Skip to content
Blog / Developer Resources

How to Give Your Hermes Agent Its Own Email Inbox

Binoy Perera

Updated

Wire AgentMail to Hermes via MCP: typed tools at startup, per-subagent inboxes, CLI setup, config.yaml, and how it differs from OpenClaw’s SKILL.md pattern-without SMTP or OAuth.

  • hermes
  • hermes agent
  • nous research
  • agentmail
  • mcp
  • model context protocol
  • agentmail mcp
  • openclaw
  • ai agents
  • email agent
  • programmatic inbox
  • hermes email
  • subagents
  • agentmail hermes

Your Hermes agent needs email for the same reasons you do: outbound outreach, catching verification codes, handling inbound replies. The default options don't work. Gmail bans automated accounts. SendGrid is send-only. Building receive-and-reply infrastructure yourself is weeks of work. AgentMail is email built specifically for agents. Create an inbox in one API call. Send and receive immediately. No SMTP config, no OAuth. Because it connects to Hermes via MCP, every subagent Hermes spawns can provision its own isolated inbox for its task, automatically, in under a second.

How this is different from OpenClaw

OpenClaw adds email through a SKILL.md file. The agent reads it at runtime and the LLM interprets it to generate API calls. That works until it doesn't. Wrong parameter, malformed request, misread endpoint. It fails silently. The email doesn't send and you don't know why.

Hermes registers AgentMail as an MCP server. At startup, every tool is registered as a typed function with a defined schema: create_inbox, send_message, reply_to_message, list_threads. The agent calls the function directly. Bad parameter means an immediate explicit error. Wrong API key means you know on startup. Nothing fails silently.

OpenClaw + SKILL.mdHermes + MCP
Tool registrationLLM interprets markdown at runtimeTyped schema registered at startup
Call executionLLM generates API callsDirect function call
Failure modeSilent. Email doesn't send.Explicit error immediately
Model sensitivityVaries by model and versionModel-agnostic
Subagent supportShared skill contextEvery subagent owns its toolset

What you need

  • Hermes installed on your machine
  • An AgentMail API key (free at console.agentmail.to, no card required)
  • Node.js 18+

Step 1: Install Hermes

Bash

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

Works on Mac, Linux, and WSL2. Takes about 3 minutes. When it finishes, the setup wizard runs automatically. Choose Quick setup, pick your model provider, and skip messaging platforms for now.

When done, reload your shell so PATH includes Hermes, or open a new terminal. Use the rc file your shell actually loads (not every install uses zsh):

Bash

source ~/.zshrc   # zsh (default on macOS)

Bash

source ~/.bashrc  # bash (common on Linux and WSL2)

If you use another shell (e.g. fish), run its equivalent reload or start a fresh session.

Step 2: Install the AgentMail CLI and confirm your key works

Bash

npm install -g agentmail-cli

Bash

export AGENTMAIL_API_KEY=YOUR_API_KEY

Bash

agentmail inboxes create --display-name "Test Agent"

You should get back:

JSON

{  "inbox_id": "boldstory348@agentmail.to",  "display_name": "Test Agent",  "created_at": "2026-04-07T00:00:00Z"}

If you see an inbox ID, your key works and you are ready for the next step.

Step 3: Add AgentMail to Hermes

Open the Hermes config:

Bash

nano ~/.hermes/config.yaml

Scroll to the bottom and add:

YAML

mcp_servers:  agentmail:    url: "https://mcp.agentmail.to/mcp"    headers:      x-api-key: "${AGENTMAIL_API_KEY}"

Save with Control + O, Enter, Control + X.

Start Hermes:

Bash

hermes

When Hermes loads, look for this line:

Code

43 tools · 76 skills · 1 MCP servers

That 1 MCP servers is AgentMail. All tools are registered and ready. If you update the config while Hermes is running:

Code

/reload-mcp

No restart needed.

Here is the full list of tools Hermes registers from AgentMail:

ToolWhat it does
create_inboxCreates a new inbox with a unique email address
list_inboxesLists all inboxes
get_inboxGets details for a specific inbox
delete_inboxDeletes an inbox
send_messageSends an email from an inbox
reply_to_messageReplies to a message in a thread
list_threadsLists threads in an inbox
get_threadGets a full thread with all its messages
get_attachmentDownloads an attachment from a message

Step 4: Your agent's first email

In the Hermes chat:

Code

Create an inbox called outreach-agent and send an emailto hello@example.com with subject "First email from Hermes"and body "Sent via AgentMail MCP."

Hermes calls create_inbox, gets the inbox ID, then calls send_message. Two typed function calls. The email lands.

To check for replies:

Code

Check the outreach-agent inbox for any new messagesand summarize what came in.

Hermes calls list_threads, reads the results, and summarizes. Your agent is now a full email correspondent.

Step 5: Every subagent gets its own inbox

This is where Hermes and AgentMail become more than the sum of their parts.

Hermes delegates tasks to isolated subagents running in parallel. Each subagent gets its own conversation context, its own terminal, and its own access to the full AgentMail MCP toolset. That means each subagent can create its own inbox for its task, use it, and delete it when done. No shared inbox state. No message bleed between agents. If a bad email lands in one subagent's inbox, it is scoped to that task alone.

Tell Hermes:

Code

Spin up three parallel subagents. Each one should:1. Create its own AgentMail inbox2. Sign up for [service] using that inbox address3. Check the inbox for a verification email and extract the code4. Complete signup with the code5. Report the account details back6. Delete its inbox

What happens:

Code

HERMES (parent)│├── SUBAGENT-1│   └── create_inbox → signs up → extracts OTP → reports → delete_inbox│├── SUBAGENT-2│   └── create_inbox → signs up → extracts OTP → reports → delete_inbox│└── SUBAGENT-3    └── create_inbox → signs up → extracts OTP → reports → delete_inbox

Three parallel tasks. Three isolated inboxes. This pattern works for signup automation, parallel outbound threads, QA testing, and any workflow that benefits from isolated email identities at scale.

Scheduled email tasks

Hermes has a built-in natural language cron scheduler. Tell it what you want and it creates the schedule.

Code

Every weekday at 9am, check the support-agent inbox forunread messages, classify each one by urgency, draft repliesfor anything urgent, and send me a summary on Telegram.

The task runs unattended through the Hermes gateway. No cron syntax, no config files.

Suggested Reading

New to agent email? Start here.

Why AI Agents Need Email The case for dedicated agent inboxes and why shared credentials break at scale.

AgentMail + Google ADK Same MCP pattern, different framework. If you are evaluating ADK alongside Hermes, read this next.

How To Give Your Browser Use Agent an Email Inbox Using AgentMail Browser automation and inbox provisioning in the same session. Signup flows, OTP handling, the full loop.

AgentMail + OpenClaw How the SKILL.md integration works if you are running both frameworks.

FAQ

Questions, answered.

AgentMail is email built specifically for AI agents, giving them real, dedicated inboxes for tasks like outbound outreach, catching verification codes, and handling inbound replies. Standard options like Gmail ban automated accounts, and services like SendGrid are send-only, making AgentMail the purpose-built alternative.

AgentMail connects to Hermes via the Model Context Protocol (MCP), which registers every email tool as a typed function with a defined schema at startup. This means the agent calls functions like create_inbox and send_message directly, and any bad parameter or wrong API key produces an immediate, explicit error rather than failing silently.

Hermes registers nine tools from AgentMail: create_inbox, list_inboxes, get_inbox, delete_inbox, send_message, reply_to_message, list_threads, get_thread, and get_attachment.

Yes. Because each subagent has access to the full AgentMail MCP toolset, it can create its own inbox for its task, use it, and delete it when done. This means no shared inbox state and no message bleed between parallel agents.

You need Hermes installed on your machine, a free AgentMail API key from console.agentmail.to (no card required), and Node.js 18 or higher.

Open ~/.hermes/config.yaml, add the AgentMail MCP server block with your API key, and start Hermes. If Hermes is already running, use /reload-mcp to register the new tools without restarting.

Yes. Hermes has a built-in natural language cron scheduler, so you can describe a recurring task in plain language and it will run unattended through the Hermes gateway with no cron syntax or config files required.

With OpenClaw, the LLM interprets a markdown file at runtime to generate API calls, which can fail silently if a parameter is wrong or an endpoint is misread. With Hermes and MCP, tools are registered as typed schemas at startup and called as direct functions, so failures surface immediately and behavior is model-agnostic.

Blog

Suggested reading

View all posts

Developer Resources · April 7, 2026

Build an Email Agent with Google ADK and AgentMail

How to wire Google's ADK into AgentMail's MCP server so your agents can send and receive email without touching SMTP or OAuth. Python setup, stdio transport, and a few things worth knowing before you ship it.

Prompt

Give yourself access to AgentMail by following the guide at:

https://docs.agentmail.to/agent-onboarding.md

and refer to:

https://docs.agentmail.to/llms.txt

for the documentation index.

Start integrating AgentMail into your AI agents today.