Agent panes

Not shipped yet

Agent panes are upcoming work. They are compiled out of every release build, and nothing on this page is something you can use today. It describes where the feature is heading, so treat the details as subject to change.

The plan is a pane that runs Claude Code or Codex over the Agent Client Protocol, with the conversation drawn natively rather than as terminal output: streaming Markdown, Mermaid diagrams, reasoning, plans, tool calls with typed payloads (a diff renders as a diff, not as JSON), live command output with exit codes, subagent timelines, and permission approvals.

Adapter versions are pinned and their cache is warmed at launch, so a pane should open without waiting on the network. Detaching and reattaching replays the conversation.

Trying it early

Agent panes need two things turned on, and both are off by default. Build with the cargo feature:

cargo build --features agent-pane

then set experimental-agent-pane = true in zz/config, or use the toggle under Settings → Advanced → System → Experimental. Without the cargo feature the config key reads false whatever the file says. While either is off, nothing creates an agent pane: not the pane picker, not the command palette, not the CLI.

Expect rough edges. This is the path for people who want to follow along, not a supported configuration.

Getting things in

Once a pane exists, you pipe into it from any shell:

git diff | zz agent-send                 # into the composer
rg -n TODO | zz agent-send --submit      # send it outright
zz agent-send --context src/main.rs:10-40 < snippet.txt

Without a target it goes to the window’s most recently focused agent pane.

prefix e sends the last command and its output from a terminal pane to the agent pane next door. It reads OSC 133 shell-integration marks to find the boundaries, so you need a shell that emits them (starship, or the ghostty, kitty and wezterm integrations all do). The capture is tail-capped, so you send the end of a failure rather than the whole build log.

Pasted images are attached inline, normalized once so the thumbnail and the bytes on the wire match. The element picker in a browser pane produces text and a screenshot as one paste.

Letting an agent drive zz

Every agent process inherits ZZ_PANE, ZZ_SESSION, and ZZ_SOCKET, and zz tools prints a short catalog written for agents. An agent can work the workspace it lives in through the same CLI you use, with no MCP server in between:

zz capture-pane -t %1 -p         # read a terminal
zz capture-browser -t %4 -o /tmp/shot.png
zz set-browser-url -t %4 https://localhost:3000
zz agent-send -t %5 'check the failing test'

zz tools works in any build. The rest of that list needs a browser pane or an agent pane to aim at.