herdr: An Agent-Native Terminal Runtime for Coding Agents

Forum topic · QianXun · 2026-08-13

Summary

herdr (github.com/herdrdev/herdr) is a Rust-based terminal runtime purpose-built for managing multiple coding agents simultaneously. Unlike tmux, which only persists processes, herdr understands agent lifecycle states (blocked, working, done, idle, unknown) via lifecycle hooks and screen-manifest pattern matching, covering 20 agents including Claude Code, Codex, and Cursor. It provides a socket API and agent skill that lets agents spawn, prompt, and rescue each other, freeing humans from per-screen babysitting. Version 0.8.0 (August 2026) cut total CPU usage by 89–95% through three targeted optimizations, documented transparently. The project ships as a single Rust binary, runs locally with no telemetry, and is licensed Apache-2.0 after a July 2026 relicense that preceded the Y Combinator F26 announcement. Windows support remains beta. DHH has begun contributing.

Overview

herdr is an agent-native terminal runtime (also called an "agent multiplexer") developed primarily by Can Celik (GitHub: Ogulcan Celik). Released as v0.1.0 on 2026-03-27, it reached v0.8.0 by 2026-08-03 and joined Y Combinator's F26 batch on 2026-08-06. As of 2026-08-13, the repository showed 28,337 stars, 2,009 forks, 79 releases, and 380,624 cumulative downloads (per GitHub API). The project is licensed Apache-2.0, after switching from AGPL-3.0 on 2026-07-22 — a relicense that explicitly preceded the YC announcement.

Core Concept

herdr treats agents as tenants and itself as the landlord. A persistent background server owns the PTY processes, workspace layouts, and pane states, while any TUI client (including herdr's own) can detach and reattach without disrupting running agents. The tagline: *"tmux is an inn for terminals; herdr is an inn with a front desk — it houses processes, understands state, and provides a phone for tenants (agents) to call each other."*

Key Points

Five-state semantics distinguish herdr from tmux: blocked (needs human approval), working (active), done (finished, unseen), idle (waiting), and unknown. State detection uses two parallel tracks: lifecycle hooks from integrated agents (Pi, OMP, Kimi, OpenCode, Kilo, MastraCode) and screen-manifest pattern matching for the rest (Claude Code, Codex, Cursor Agent CLI, Copilot, Devin, Grok, Qwen Code).
Detection is conservative by design: unrecognized UI falls back to idle, never triggering input or destructive actions. Manifests are hot-updatable from herdr.dev but can be disabled locally; local ~/.config/herdr/agent-detection/<agent>.toml files always override.
Agent-to-agent control plane: a Unix socket / Windows named pipe API plus a bundled agent skill (herdr --skill) lets agents split panes, spawn peers, prompt with --wait, and rescue stalled peers. A 5-second stall detection returns agent_prompt_stalled rather than hanging.
Persistence tiers span detach/reattach (full session preserved), server restart (layout + optional pane history), and update --handoff (experimental live migration). Pane history replay defaults to off to avoid retaining secrets in session files.
Performance work in 0.8.0 delivered 89–95% total CPU reduction by removing redundant animations, suppressing render fan-out from hidden panes, and coalescing mouse-move repaints. Examples: 1 silent working pane dropped from 1.467% to 0.133% CPU; 10 panes (9 hidden, 3 clients) dropped from 13.014% to 0.617%. The accompanying blog post is unusually candid, including scenarios where optimization gains shrink to ~23.6% when PTY read and parser work dominate.
Local-first architecture: no account, no telemetry, no cloud sync. Sessions live as plain JSON under ~/.config/herdr. The only network contact is an optional manifest update check.
Single Rust binary (~22.4k lines, 245 .rs files) using ratatui, crossterm, and a patched portable-pty. No Electron. Linux/macOS are stable; Windows is beta with bundled ConPTY runtime.
Re-licensing posture: the AGPL → Apache-2.0 change on 2026-07-22 preceded the F26 announcement by two weeks, framed as a public commitment to keep the runtime free. Commercial plans target multi-machine orchestration and non-TUI clients — not the runtime itself.
Plugin ecosystem (~500+ plugins within a month of marketplace launch) uses any executable; the full herdr CLI doubles as the SDK. Plugins run with the user's permissions — no sandboxing — and herdr plugin install previews manifest and scripts before confirmation.
Notable contributor: DHH began submitting PRs on 2026-08-09 (tab centering, tab-move shortcuts, configurable pane borders, window-title sync), a strong vote of confidence from the Rails creator.
Honest limits: screen-manifest detection is brittle against upstream UI changes (Claude Code's switch from Braille dots to half-circle spinner in 2.1.228 required a herdr patch); remote manifests are a trust surface even with SHA-256 verification; Windows remains a preview; multi-machine orchestration is still aspirational; the core team is still essentially one person transitioning into a startup.

Verdict

herdr addresses a real gap in the 2026 multi-agent landscape: agents are proliferating, but no one has built solid tooling to *herd* them. Its pitch is deliberately classical — no agent of its own, just the runtime that hosts yours — backed by strict architectural rules (state/runtime separation, pure render functions, multiplicative-cost accounting), transparent engineering writing, local-first operation, and an Apache-2.0 license preserved through YC funding. Worth installing for anyone juggling three or more coding agents daily.