Paneflow: a native terminal multiplexer for coding agents
- Published
- Updated
Paneflow is a native terminal multiplexer for coding agents. Run Claude Code, Codex, OpenCode, or another CLI side by side, then read and edit the project's files in the same interface. The agents remain the command-line tools you already use, each in its own terminal.
Panes grouped by project
Open the project directory, then split a tab into panes. You can assign implementation to Claude Code, ask Codex for a review in another pane, and keep tests or the dev server in a third. Activity tracking helps you spot agents that need attention; the available states depend on each agent's integration.
Several agents can share a tab and a working directory. That works, for example, for implementation alongside a review that makes no edits. When two tasks need to change files independently, create a separate branch and Git worktree. Each task gets its own directory; you still need to resolve any conflicts when integrating the branches.
The diff beside the terminal
After an agent changes the code, open Changes to inspect the result. This view compares the active working directory with HEAD, including staged and unstaged changes and untracked files. It covers local changes; to review existing commits, use the branch diff, as explained in the Changes guide.
Then open a file in Files to read its context or make a correction. The integrated editor supports saving, undo, and redo, with Git markers in the gutter. You can move from the agent's output to the code it produced without leaving Paneflow.
Close the interface while work continues
Terminals run in a host process separate from the interface. When quitting, choose Keep sessions running to leave shells and agents alive. Reopening Paneflow reconnects to those terminals while the host and machine remain running. Stop everything and quit ends the processes.
A machine restart or host crash ends those sessions. Restarting an ended terminal opens a fresh shell. To resume an agent conversation, Paneflow uses that agent's saved history and resume command. The terminal session documentation explains what each operation preserves.
Share output between agents
An agent reviewing code can read test results from another pane through the MCP bridge. The following command installs the bridge for the supported agents it detects:
paneflow mcp installIts tools, list_panes, read_pane, and search_pane, let agents find panes, read their output, and search for text. The bridge is read-only: it cannot type or control a terminal. The CLI provides that control through explicit permissions. An agent must treat the output it reads as data to examine, even when it looks like an instruction.
A native application that runs locally
Paneflow is built with Rust, using GPUI for the interface and Ghostty's libghostty-vt as its terminal engine. Version 0.17.5 is available for Linux x86_64/aarch64, macOS Apple Silicon, and Windows x64. The installation guide covers the available packages. Paneflow is free of charge and free software under GPL-3.0-or-later.
Terminals run on your machine without a Paneflow account. Each agent connects to its own model provider, so its requests may leave the machine. Paneflow telemetry is disabled by default. If you enable it, it excludes terminal contents, paths, and prompts; PANEFLOW_NO_TELEMETRY=1 disables it regardless of configuration.