Herdr Complete Guide 2026: The New Standard Runtime Where Any Agents Can Talk to Each Other

「I want to run Claude Code and Codex at the same time and have them work together…」「Running multiple agents in separate terminals, I lose track of who's stuck…」

The answer to these problems is the hot open-source project of 2026, Herdr. It's an agent runtime that lets you run 19 kinds of AI coding agents — Claude Code, Codex, Cursor, Grok, Hermes, and more — all in one terminal, and have agents talk and collaborate with each other.

On GitHub it has gathered 25,000+ stars, 360,000+ total installs, and 530+ community plugins. It's backed by Y Combinator.

This guide fully explains Herdr's mechanism, agent-to-agent conversation, setup, and usage for beginners.

What you'll learn in this article

What Herdr is (why it's called an "agent runtime")
The mechanism that lets agents talk to each other
Setup (one command)
Real usage (CLI and automation commands)
Supported agents and cautions

What is Herdr?

Herdr is a runtime where AI coding agents "live."

Traditionally, running multiple agents at once meant opening and managing separate terminals. Herdr centralizes them all on one server.

Key features

Always running: a background server. Even if you close the lid or lose network, agents keep working
Agents can talk to each other: via CLI and socket API, an agent can send prompts to another agent
Automatic state detection: shows each agent's status — working / blocked / idle
Run existing agents as-is: auto-detects 19 kinds including Claude Code, Codex, Cursor, opencode, Grok, Copilot, Hermes

Why "agent-to-agent conversation" is impressive

Herdr's biggest feature: agents can talk to each other.

Mechanism (diagram)

Setup (one command)

curl -fsSL https://herdr.dev/install | bash

That's it. Herdr detects installed agents automatically.

Usage

Start Herdr

herdr

The server starts, and agents you have installed (Claude Code, Codex, Cursor, etc.) are detected automatically.

Give another agent a job

agent prompt B "Refactor the auth module and report back"

Wait for completion

agent wait B

Waits until Agent B finishes (or becomes blocked).

Check status

The sidebar shows working / blocked / idle for all agents at a glance.

Supported agents (19 kinds)

Claude Code, Codex, Cursor, opencode, Grok, Copilot, Hermes, and more — 19 kinds in total. Community plugins: 530+.

Cautions and limitations

blocked detection is carefully designed: with the screen-manifest method, it only marks blocked when a known approval/question/permission UI is matched (prevents false positives)
Approvals are human-only: messages from other agents never replace human approval
Some features are still early: while a new agent's UI isn't supported, it may show idle instead of blocked
One session per agent: physically, one pane = one agent as a rule

Summary

Herdr is a new foundation for "housing multiple AI coding agents in one runtime and letting them talk to each other."

Agent-to-agent conversation: agent prompt assigns work to another agent, agent wait waits for completion
Always running: the server works in the background; work continues even when the lid is closed
Status visibility: working / blocked / idle at a glance in the sidebar
19 agents supported: use existing agents as-is (Claude Code, Codex, Cursor, Grok, Hermes, etc.)
Y Combinator-backed and open source (Apache 2.0 / GitHub: herdrdev/herdr)

For anyone who wants to "divide work among agents" or "manage multiple agents' parallel work on one screen," this is the tool you've been waiting for. Install it with one command, assign different tasks to two agents, and experience the future where agents talk to each other.