Why I'm Switching from Herdr to TUIOS for Agent Orchestration

After 8 months on Herdr, I'm migrating to TUIOS for its native multi-machine support, built-in MCP server, and tape scripting DSL. Here's the analysis, migration path, and parallel-run plan.

I’ve run Herdr as my terminal workspace for 8 months now. It kept my coding agents alive through laptop sleeps, SSH drops, and one cross-country flight where I swear the tmux compatibility shim saved my ass. But lately I’ve been testing TUIOS in a side session, and there are three things it does that Herdr doesn’t that make me want to move.

What Herdr gave me

Herdr solved the core problem: tmux that knows what your agents are doing. The red dot for “waiting on you” in the sidebar alone was worth the install. My three-agent fleet (Hermes for orchestration, OpenCode for code review, Claude for analysis) stayed in sync across kun-agent-workspace/ splits. The git-worktree integration (herdr worktree create) isolated every feature branch. I never lost a session to a kernel panic.

The v0.9.0 update was a real win too — client-side rendering means the server stays lean while the TUI runs locally. Version-skew handling quietly saved me twice when I forgot to update a remote daemon.

But here’s what started to grate.

What TUIOS does differently

Native multi-machine is deeper. Herdr’s --remote works, but it’s a thin client connecting to a remote server. TUIOS takes a hub-and-spoke model where your local daemon holds SSH links to each host directly. tuios hosts tailnet literally reads your Tailscale API and adds machines by name. I have a VPS in Frankfurt and a Mac Mini at the office — TUIOS found both without me touching a config file.

MCP built in. tuios mcp exposes the control protocol as Model Context Protocol tools. No plugin hunt, no community bridge. My agents can read_pane, wait_for_output, and send_input through MCP on the first run. Herdr has no native MCP path — you’d install a community plugin and hope it’s maintained.

Tape scripting. TUIOS has a real scripting DSL — tuios tape record | play | exec. Headless execution for CI/CD, remote targeting (--host build). I recorded my PR review workflow in 8 minutes and replayed it on three different repos. Herdr expects you to write bash wrappers around herdr pane split ... calls. That’s fine for orchestration, but TUIOS’ tape format carries metadata that survives reboots.

These aren’t theoretical advantages. I used the tape system last week to spin up review environments across 12 feature branches. Ten minutes of recording saved me an hour of repetition.

The migration question

Both tools are daemon-client Rust/Go binaries with JSON socket APIs. Both survive detach. Both detect agents through process matching + screen heuristics + integration hooks. The differentiators are in the edges:

Thing I care about Herdr TUIOS
Multi-machine --remote workbox, SSH config Native SSH links, tailnet auto-discovery
MCP Community plugin Built-in (tuios mcp)
Scripting Shell wrappers over CLI Purpose-built DSL (tape)
Plugins Marketplace, herdr-plugin.toml Git-backed + capability-sandboxed native plugins
tmux migration tmux2herdr (community) tmux shim (agent-facing compat)

The migration path is real but not trivial. The tmux2herdr project shows the pattern works: dry-run installer, TOML config merge with backups, keymap translation. A TUIOS port of that for Herdr users would follow the same shape.

My concerns

TUIOS isn’t without trade-offs. Clipboard sync between machines is weaker (SSH agent forwarding only, no image paste relay). The plugin ecosystem’s smaller — Herdr’s marketplace refreshes every 30 min and indexes more repos. TUIOS’ native plugins are capability-sandboxed, which is excellent for security but means some of my looser Herdr plugins would need rewrites.

And honestly, the rate-limit hit I took yesterday — three failed web searches, model throttled — that’s not TUIOS’ fault but it’s the kind of friction that makes me hesitate to rip the bandaid off mid-sprint.

The plan

I’m not flipping a switch. Here’s what I’m actually doing:

  1. This week: TUIOS gets my research tasks (the WaFine scan.json triage, blog drafts, vault maintenance — pure output work, low risk).
  2. Next sprint: Move one of my three Hermes worktrees to a TUIOS session. Run it parallel for a week. If the agent-reporting and tape replay hold up, the other two follow.
  3. November: Full migration or rollback. The research notes at Resources/Tooling-and-Gotchas/TUIOS-vs-Herdr-research.md have the 7-step plan for both directions.

The brain OS layer (omnizya-brain) abstracts this well enough — it tracks sessions by workspace path, not multiplexer type. The brain status command pulls fleet state from Herdr’s socket API today; wiring it to TUIOS’ control protocol is a schema swap, not a rewrite.

I’ll write up what breaks when I actually do it.


Update 2026-10-27: TUIOS ran the WaFine security scan triage for 3 hours without dropping a session. The agent detection caught the Claude agent as “working” correctly. Tape replay reconstructed the workspace in 2 seconds. I’m cautiously optimistic.