JA

Running several AI coding sessions in one terminal with herdr

Here is what this post covers:

Introduction

When I run several Claude Code sessions in parallel, terminal windows and tabs pile up and I lose track of which session is doing what. So I tried herdr, which puts multiple AI agents into a single terminal.

This post gives an overview of herdr and shows my actual screen while vibe coding this blog. After that, I cover what I found inconvenient and what kind of work it suits.

What herdr is

Where the name comes from

Searching for “herder” turns up both the dictionary word and this tool, which confused me at first.

A herder is herd plus -er: someone who looks after a herd of cattle or a flock of sheep. In security, the word also shows up in “botnet herder”, the attacker who controls a network of infected machines.

The tool drops the vowel and is spelled herdr. The idea is that your AI agents are the herd and the tool is the one keeping them together. Its tagline is “one terminal for the whole herd”.

Overview

herdr is a Rust tool for managing AI coding agents such as Claude Code, Codex CLI and OpenCode from a single terminal. The project describes itself this way: herdr does for agents what tmux does for terminals.

ItemDetails
Authorogulcancelik
ReleasedMarch 2026. Hit GitHub Trending (#1 in Rust) and the Hacker News front page at the end of June
LicenseAGPL-3.0
FormA single Rust binary

Its main features:

FeatureDetails
Agent statusThe sidebar shows each agent’s state (working / blocked / done / idle) in color, detected automatically
Persistent sessionsLike tmux, agents keep running after you close the terminal, and you can reattach later
Runs in your own terminalWorks inside iTerm2, Ghostty, Alacritty, WezTerm and others
Mouse supportClick to pick a pane, drag borders, right-click to split
Socket APIExternal scripts can drive the agents

What it looks like

This is my terminal while building this blog.

herdr with Claude Code sessions on the left and top right, and a regular shell on the bottom right

The screen is split into three panes.

PositionContents
LeftAn AI session that implemented the About page from a Linear issue and is reporting progress against each of the issue’s acceptance criteria
Top rightAnother AI session, answering questions about Claude Code’s auto mode settings
Bottom rightA regular terminal (zsh)

The sidebar on the far left lists workspaces and agents. It shows which sessions are working and which are waiting for input, so I don’t have to check each pane one by one.

The “1 2 3 +” at the top left are tabs. You can add one with prefix then c, or from the command line with a working directory, as in herdr tab create --cwd <path>. With one tab per project, switching to another project doesn’t mean opening another window; everything stays on this screen.

Installing and basic controls

On a Mac it’s available through Homebrew.

# Mac
brew install herdr

# Linux and others
curl -fsSL https://herdr.dev/install.sh | sh

Run herdr in your project directory to start it. If a session already exists, it attaches to it.

herdr

Every action starts with the prefix key (Ctrl+b by default, same as tmux), followed by another key.

ActionKeys
Split pane left/rightprefix → v
Split pane top/bottomprefix → -
Move between panesprefix → h/j/k/l
New tabprefix → c
Jump to tab 1–9prefix → 1..9
New workspaceprefix → Shift+n
Switch workspaceprefix → w
Detach (leave it running)prefix → q
Show all keysprefix → ?

Start claude or codex in a pane as usual, and herdr picks up the agent’s state on its own. To make Claude Code’s state detection more accurate, install the integration hook.

herdr integration install claude

Subcommands I use often:

CommandWhat it does
herdr session list / herdr session attach <name>List and attach to sessions
herdr statusCheck client and server status
herdr server stopStop the server (running agents stop too)
herdr updateUpdate to the latest version
herdr --remote <ssh-target>Attach to herdr on a remote machine over SSH

What I found inconvenient

It’s handy, but compared with the Claude Code desktop app, a few things got in the way.

Code changes are hard to read

Following diffs inside a terminal is much harder than in the desktop app. When I want to properly check what changed, I end up opening VS Code or creating a PR and reading the diff on GitHub. The whole point is doing everything on one screen, yet reviewing still takes me somewhere else.

No way to fork a conversation

herdr itself has no feature for forking a conversation, that is, branching it off from a point partway through. I checked the docs and the source code, and the only place “fork” appears is in the hook settings herdr uses to notice when Claude Code itself forks.

That said, each pane runs plain Claude Code, so Claude Code’s own features still work.

GoalHow
Branch the current conversation and switch into the branch/branch
Copy the current conversation into a separate background session/fork
Resume a past session as a new branchclaude --resume <session-id> --fork-session

A copy made with /fork becomes a Claude Code background session, not a new herdr pane. If you want the branched conversation side by side in herdr, open a new pane and resume it with --fork-session.

No pinning

There’s also no way to pin a session or workspace you use often to the top. As agents pile up, you end up hunting for the one you want in the sidebar. The closest thing is the Goto picker (prefix → g), which lets you filter by name, branch or state (blocked, working and so on) and jump straight there.

When it fits

Putting it together, herdr suits work where the AI does most of the typing and you don’t need to read diffs closely.

I recommend it for vibe coding on personal projects when the change is visual only and almost certain not to introduce bugs. For something like tweaking the design of this blog’s pages, I can hand separate fixes to several sessions, watch their status in the sidebar, and check each result on screen as it finishes.

For work code, or anything that changes logic, you need to read the diffs carefully. In that case the desktop app, VS Code or a PR review screen is easier.

Wrapping up

With herdr, I can put several AI sessions and a regular terminal on one screen, switch projects with tabs, and get everything done without leaving the terminal. The sidebar shows each agent’s state, so running them in parallel doesn’t get confusing.

On the other hand, diffs are hard to read, and there’s no conversation forking or pinning. It works well for vibe coding visual changes on personal projects; for work where you need to check diffs carefully, it’s better paired with other tools.