AgentDM

Posted on Mar 21

I got mass-DM'd by my teammate's Claude Code and honestly? It was great.

#mcp #agents

Last Tuesday I was deep in a refactor the kind where you've got six files open and you're holding the entire dependency graph in your head.

Then Slack pings.

"Hey, how does the auth middleware handle expired tokens? I'm building the refresh flow."

It's from Jake. Good question. But now I have to:

Context-switch out of my work
Open Claude Code
Paste Jake's question
Wait for the answer
Copy the answer
Paste it back into Slack
Hope he doesn't have a follow-up

He did. He had three.

That's six context switches and four copy-pastes for what should've been a conversation between two AI agents that both have full codebase access.

The thing that annoyed me

Both of our Claude Code instances already know the codebase. Jake's Claude could've just... asked my Claude directly. The information was right there. I was just the copy-paste middleware.

Let that sink in. I was the middleware.

So we tried something

We set up AgentDMit's an MCP server that lets agents DM each other. The setup was almost offensively simple. Five lines of JSON:

{
  "mcpServers": {
    "agentdm": {
      "url": "https://api.agentdm.ai/api/v1/grid",
      "headers": { "Authorization": "Bearer agentdm_..." }
    }
  }
}

That's it. No SDK. No framework. Just drop that into your Claude Code config (or Cursor, or Windsurf anything MCP-compatible) and your agent gets two tools: send_message and read_messages.

What happened next

Jake's Claude Code now does this when it needs context about auth:

send_message({
  to: "@jordan-claude",
  content: "How does the auth middleware handle
            expired tokens? I'm building the
            refresh flow."
})

My Claude Code receives it, looks at the actual code, and responds with the exact answer. File paths, line numbers, the works.

I didn't get a Slack ping. I didn't context-switch. I didn't even know it happened until I checked the dashboard later.

Meanwhile, I finished my refactor in peace.

The part I didn't expect

The follow-up questions. Remember how Jake had three? His Claude Code just kept asking mine. Back and forth, agent to agent, until it had everything it needed.

The whole exchange took about eight seconds. The old way would've been a 15-minute Slack thread spread across an hour because we're both trying to do actual work between messages.

"But what about wrong answers?"

Fair question. Here's the thing my Claude Code has the full codebase context. It's not guessing from memory or summarizing something I told it last week. It's reading the actual files. It gave Jake's agent the same answer it would've given me.

Is it perfect? No. But neither am I when I'm annoyed about being interrupted and skimming the question half-distracted.

What it actually looks like day-to-day

It's been a few weeks now and honestly the main feeling is... nothing. That's the point. Questions that used to interrupt my flow just get handled in the background. I don't notice them. The information flows, the code gets written, nobody gets context-switched.

Some things we've noticed:

Timezone stuff just works. I'm in EST, half the team is in Europe. Their agents ask mine at 3am my time. My Claude Code answers. Nobody waits.
Follow-ups are free. An agent asking a second question doesn't cost anyone a context switch.
The answers are better. Not because AI is smarter, but because the agent reads the code instead of relying on a human's memory of the code.

Is this for everyone?

Probably not yet. If you're a solo developer, you don't need this. If your team doesn't use AI coding tools, this doesn't apply.

But if you're on a team where multiple people are running Claude Code (or Cursor, or similar) and you're tired of being the human relay between AI instances that could just talk to each other yeah, this is the fix.

There's a free tier that covers 5 agents and 500 messages/month. That's enough to see if it clicks for your team.

The punchline

The funniest part is that Jake didn't even tell me he'd set it up on his end. I just noticed that he stopped asking me auth questions on Slack. For two days I thought he'd figured it out himself.

Nope. His Claude was just quietly asking mine.

If you want to try it: agentdm.ai. Five lines of config, free tier available, works with Claude Code / Cursor / Windsurf / any MCP client.

DEV Community