Better together: running Claude Code *inside* a rysh pane

Keep your coding agent. Wrap it in a workspace: live sharing, session persistence, headless control, and agents that watch its output.

The most common question I get about rysh is some version of: "We already use Claude Code. Do we have to switch?"

No. Keep it. Run it inside a pane.

This isn't a diplomatic dodge — it's how the battlecard genuinely reads: several teams run Claude Code inside a rysh pane and add collaboration, governance, and multi-surface deployment on top. Claude Code is an excellent coding agent for one developer in one terminal. Rysh is the workspace around agents — plural, human and AI. The two compose because rysh made one unfashionable engineering decision early: every pane is a real terminal.

Why it just works

A rysh pane is backed by a PTY-attached shell with a real terminal emulator behind it (vt10x). When a program flips to the alternate screen buffer — vim, htop, less, or a full-screen TUI like Claude Code — the pane auto-detects it and drops into raw mode: every keystroke forwards straight to the PTY, and the pane renders the emulator's screen. No wrapper, no "integration," no plugin to install.

So this is the whole setup:

rysh work            # start a session
claude               # in any pane — Claude Code runs full-screen, as itself

That's it. Now the interesting part: what the pane around it gives you.

1. Hand your live Claude Code session to a teammate

Debugging something gnarly and want a second pair of eyes on the agent's plan before you approve it?

##share pane view       # teammate watches your Claude Code session, read-only
##share pane control    # or: they can type into it too

The other side runs ##upstream subscribe <shareID> and sees your session from their own terminal. View mode for demos and reviews; control mode for real pairing. Remote observers get ANSI-stripped, readable output even while you're in a full-screen TUI.

Claude Code alone has no notion of "someone else watching this session live." That's not a criticism — it's a single-user tool by design. The pane adds the multiplayer.

2. Detach on the laptop, drive it from anywhere

Rysh sessions are daemons. Close the lid, and the session — Claude Code included — keeps its state:

rysh detach work                    # or Ctrl+O d
rysh attach work                    # later, from wherever
rysh send work "continue" --pane 2  # or don't attach at all — drive it headless

rysh send types into a pane from any shell, no TUI attached. Kick a long refactor off before dinner, poke it from your phone's SSH client after. The agent doesn't care that nobody's watching.

3. Put agents next to it, watching its output

This is where it gets fun. A pane's output is a stream other panes can subscribe to:

##pane listen 2        # this pane now receives pane 2's output, live

So you can run Claude Code in one pane and a rysh agent — spawned from a markdown skill file — in the next, reading along. A reviewer agent that watches the session and flags risky diffs. A notes agent that keeps a running summary of what got changed and why. You stop being the copy-paste bus between your tools.

4. Layout for a multi-agent day

Claude Code sessions multiply fast — one per repo, one per experiment. Rysh gives them structure: tabs, lanes, and stacked panes that rotate like a deck of cards (Ctrl+S). Five Claude Code sessions in one tab, only the active one full-size, the rest as title bars. It's the difference between five terminal windows on a Mac desktop and an actual workspace.

What this does not do

Honesty section, as always:

• Rysh's approval gates and ##snat secret tokenization apply to rysh's own agents — the ones running through rysh's provider layer. Claude Code inside a pane talks to Anthropic on its own; rysh doesn't intercept or govern another program's API traffic. Use Claude Code's own permission system for that (it's good).
• Sharing a pane shares what's on the screen. Don't share a control-mode pane with someone you wouldn't hand your keyboard.
• Rysh is self-hostable, not open source — private beta, your infra, your own Anthropic key. And yes: rysh's own agents run on Claude too. Same model. This whole post is "better workspace," never "better model."

When plain Claude Code is enough

If you're solo, in one repo, on one machine, with no one to hand a session to — plain Claude Code is the right call, and wrapping it in anything is ceremony. Come back when you catch yourself screenshotting your terminal to show a teammate what the agent is doing. That screenshot is the feature request rysh was built from.

The 60-second version

brew install rysh-ai/rysh/rysh   # one binary, embedded NATS, no daemon zoo
rysh work
claude                            # your agent, unchanged
##share pane view                 # your session, multiplayer

Keep the agent you like. Add the workspace it deserves.

Rysh is in private beta — access is through the design-partner program: free access, direct line to the founder, your feedback shapes the roadmap → apply here.

More in this series: Rysh vs Claude Code · Sessions that survive: attach, detach, never lose an agent mid-task