herdr

herdr is for the person who runs long agent jobs over SSH and loses them to a dropped connection. It is a background runtime that owns the terminals your agents live in, so Claude Code, Codex, Cursor, OpenCode or Grok keep going when you detach, and you reattach later to find the session where you left it. The trait that decides it is persistence, nothing more.

It does not make branches or merge anything. Pick it if disconnects are your problem and you have two or more agents running at once. Pick Claude Squad if you want worktrees, and Superset if you want a window.

The risk is agent-to-agent prompting. The README's feature is that agents can spawn panes and prompt each other, and the runtime provides no container layer and no per-agent worktree, so an agent that instructs a second agent does so on the same checkout with the same user. Two agents editing one tree is a merge conflict without a merge tool.

Give each agent its own clone before you let them talk. What it does right: it does not wrap or replace the agents, so an agent update cannot break the runtime, and a runtime bug cannot corrupt an edit.

The design is a terminal server with agent-aware state. 1. A single Rust binary, no Electron, holds the sessions and survives disconnects and, per the README, machine restarts. 2. Every pane carries a status, so a supervisor can read whether an agent is working, waiting or done without parsing its screen. 3. A wait call blocks until another agent is genuinely blocked, which is the primitive that turns polling into scheduling.

No benchmark is published, and there is no task to benchmark; it moves no code. The observation: status tracking on a pane is the first honest attempt at observability for terminal agents.

The commercial surface is one email address: enterprise partnerships via hey@herdr.dev, no pricing, no plan and no named investors on the site. Thirty-five thousand stars for a terminal runtime says the pain is real, and a runtime that owns the terminals of every agent vendor is a neutral position worth something to a company that sells terminals.

Moat: none beyond being first; the primitives are tmux-shaped and reproducible. Likely acquirer: Warp, or an agent vendor that wants its sessions to survive a laptop lid. Position: use it, and do not be surprised by a logo change.

The demo is an agent still running the next morning. Procurement: that is also the security finding, because sixty engineers means sixty hosts holding live agent sessions with repository access after the person has left the desk, and there is no admin console, no SSO and no log of who reattached. The Windows install runs powershell -ExecutionPolicy Bypass, which our endpoint team will refuse, and the README says endpoint-protected Windows needs beta access.

It is headless and could sit in CI. Onboarding is brew. Not yet; an idle timeout and a central kill switch first.

Apache-2.0 and a single binary, installed with brew install herdr or mise use -g herdr, no runtime to babysit. The socket API is the part I care about: my own scripts can open panes, send prompts and read status without going through a vendor's CLI, which makes the runtime a building block rather than a product.

There is no MCP and no model configuration, correctly, since it never talks to a model; the agents in the panes bring their own keys and their own local endpoints. A fork is a cargo build. Small, unopinionated, and it stays out of my way.