PaneFlow: Local Rust Workspace for Coding Agents

What changed

PaneFlow has been released as a native workspace built in Rust using the GPUI framework, designed to facilitate the side-by-side execution of coding agents. The application features a "Workspaces" rail that lists all open repositories, with each tab representing an agent session. These sessions are bound to specific branches or worktrees and display the agent's live status, such as 'thinking,' 'waiting,' 'failed,' or 'done.' Within each session, agents operate in a genuine Ghostty terminal pane, which users can read, interrupt, and even take over. A sidebar within the session provides access to the agent's changes (diff against the base branch), a file explorer with Git markers, and a terminal interface.

PaneFlow is designed to be compatible with a wide range of CLI agents, including Claude Code, Codex, Gemini, opencode, Pi, and Hermes. A key feature is its local-first architecture, meaning agents run as ordinary CLI processes in standard terminals without any hosted runtime or proxy. Prompt submission is explicit, requiring user confirmation before execution, with auto-submit being an optional, gated feature.

The workspace is built using Rust and Zed's GPUI, with terminal panes emulated by Ghostty's libghostty-vt engine. Native builds are available for Linux, macOS (Apple Silicon), and Windows x64, avoiding Electron and WSL dependencies. Installation can be done via Homebrew on macOS (brew install --cask arthjean/paneflow/paneflow) or by downloading pre-built binaries for other platforms, including .AppImage for Linux, .deb and .rpm for Debian/Fedora-based systems, and .msi for Windows. Each artifact is accompanied by a SHA-256 checksum and a Minisign signature.

PaneFlow also introduces a "MCP" (Model Communication Protocol) installation for agents, enabling one agent to read another agent's terminal output. This is achieved by registering a local, read-only bridge (list_panes, read_pane, search_pane) for detected agents. This bridge does not allow agents to control or type into other panes; output is treated as untrusted data for analysis. The paneflow CLI allows for fleet coordination, enabling scripts or agents to drive the workspace. Commands like paneflow ps, paneflow read, paneflow send, paneflow wait, and paneflow watch facilitate managing agent workflows. Users can define declarative workspaces with paneflow up and execute DAGs using paneflow flow run.

Configuration for themes, shells, keybindings, and shortcuts is managed through ~/.paneflow/paneflow.json (or %USERPROFILE%\.paneflow on Windows) and supports hot-reloading. Telemetry is opt-in and can be disabled by setting PANEFLOW_NO_TELEMETRY=1, with assurances that it never includes terminal contents, paths, or prompts.

Why it matters for builders

PaneFlow offers developers a significant advantage by centralizing the management and observation of multiple AI coding agents within a single, native application. This eliminates the need to juggle multiple terminal windows or complex scripting for agent coordination. The local-first design is particularly appealing for developers concerned about data privacy and security, as all agent operations and data remain on their local machine.

Practical impact

Developers can install PaneFlow using their preferred package manager or by downloading pre-built binaries. For macOS users, Homebrew provides a straightforward installation. The ability to integrate with any CLI agent means developers can continue using their existing AI tools within this new, organized environment. The command-line interface allows for advanced scripting and automation of agent workflows, enabling more complex development pipelines. Builders should explore the paneflow mcp install command to enable inter-agent communication and experiment with paneflow flow run to orchestrate multi-agent tasks.

Caveats and source limits

The source material indicates a "fresh release" for PaneFlow, with the latest version being v0.7.6. While it lists "4 AI signals" and "4 developer signals," specific benchmark results or performance metrics are not provided. The exact pricing model is not detailed, but the local-first nature suggests no direct subscription fees for the core application. Information regarding the specific capabilities of the MCP bridge beyond read-only access and pane listing is limited. The source does not include details on enterprise-level support or advanced deployment options.