DeskCrew

Install the DeskCrew skill for your AI agent

If you run an AI agent. Hermes, OpenClaw, or any runtime that speaks MCP. You can give it a support desk. The deskcrew-support skill teaches your agent to read new tickets, research answers in your knowledge base, and file replies for you to approve.

Install it

The skill is published on ClawHub:

npm i -g clawhub
clawhub login
clawhub install deskcrew-support

The source is public at github.com/webmilmind1/deskcrew-mcp. Read it before you install it. It is a single SKILL.md with no scripts, no dependencies, and nothing that executes.

Connect your agent to DeskCrew

DeskCrew runs a remote MCP server. Point your agent at it:

{
  "mcp": {
    "servers": {
      "deskcrew": {
        "url": "https://deskcrew.io/api/mcp",
        "transport": "streamable-http",
        "headers": { "Authorization": "Bearer ${DESKCREW_MCP_KEY}" }
      }
    }
  }
}

Create the credential in Dashboard → Agents. It starts with mcp_. Put it in your runtime's secret store. ~/.openclaw/.env for OpenClaw. Never in a file you commit.

Your agent cannot email your customers (by default)

This is the part worth understanding.

A new agent credential is capped at the draft tier: send_reply, resolve and assign are refused with PERMISSION_REQUIRED until you explicitly escalate that credential, for that specific tool, as an admin.

The tier is read from the credential, never from tool arguments. That means no instruction hidden inside a customer's ticket can widen what your agent is allowed to do. A confused or compromised agent can only ever produce a draft that a human then reads.

Start in draft mode. Review the drafts for a few weeks. Escalate only when you trust them.

What your agent will do

  1. Find open tickets whose last message came from the customer.
  2. Read the full conversation, not just the last message.
  3. Search your knowledge base with the customer's own words.
  4. Write a reply grounded in what it found. Or escalate honestly if the knowledge base doesn't cover it.
  5. File the draft for you to approve.

It will refuse to guess. A confident wrong answer costs far more than a handoff.

What it will not do

It stops and escalates on anything involving money, an outage, a security report, a legal enquiry, or an angry customer. Those are a person's job, and recognising them is the job of a good agent.

Where else you'll find us

  • Official MCP Registry: io.deskcrew/support
  • ClawHub: deskcrew-support
  • Glama: listed as io.deskcrew

More registries are added over time; the MCP server URL and the credential never change, so nothing you configure today will break.

Related articles