Rysh – an AI harness where Claude and Codex agents work as a team (Go)

Rysh-ai launched the Rysh CLI, an open-source terminal multiplexer written in Go under the Apache-2.0 license. The tool delivers traditional terminal multiplexing features including tabs, lanes, split panes, and compatibility with terminal interfaces like vim and htop, while integrating AI agents directly into every pane. According to github.com's report, the repository holds the standalone command-line application, built to work as an agentic terminal harness where developers can run interactive sessions with models like Claude and Codex side by side.

This article details the technical architecture, installation pathways, execution models, and multi-agent coordination capabilities of the Rysh CLI. It is intended for software engineers, DevOps practitioners, and systems developers looking to manage parallel agentic workflows within local terminal sessions.

Rysh – an AI harness where Claude and Codex: what actually changed

The Rysh CLI transforms standard terminal panes into persistent, stateful agent environments that retain context across session restarts. Standard multiplexers kill background processes or lose unhandled state when closed, but a Rysh session runs as a persistent daemon process. Developers can assign specific tasks to individual panes using commands like ##claude or ##codex to spawn interactive vendor-specific sessions. When a session is stopped outright and restarted, the underlying daemon restores the active agents without requiring manual prompt re-entry, allowing models to resume work with prior instructions intact.

The release also introduces built-in privacy protection via a feature called SecretNAT, enabled by default. SecretNAT intercepts request bodies before they leave the host machine, replacing sensitive credentials and secrets with temporary tokens. If an API response returns a plaintext credential, Rysh flags it inside the pane without rewriting the response body. Developers can inspect or reverse tokens locally using the snat get command, ensuring raw API keys and secrets are never exposed to external model providers.

Rysh – an AI harness where Claude and Codex: how it works

Navigation within Rysh uses a multi-layered visual structure consisting of tabs, lanes, stacks, and panes. Panes stacked within a lane can be expanded or collapsed to a title bar displaying the current pane index. Each pane supports four input modes: shell mode for standard PTY commands, prompt mode for direct LLM queries, rysh mode for multiplexer controls, and chat mode for pure conversation without tool calling. Switch operations occur via double-pressing the Escape key, while complex pane grids such as three lanes across four panes can be constructed dynamically using grid allocation commands.

Behind the scenes, Rysh requires Go 1.25.3 or newer to build from source via the go install command, placing the open-source rysh binary in the Go environment path. The open-source repository carries no replace directives and builds standalone, while developing alongside its shared module requires the sibling rysh-cli-parent workspace using Make commands. An official installation script is also provided to fetch prebuilt release binaries directly from package distributions.

Rysh – an AI harness where Claude and Codex: why it matters now

Local agent coordination requires structured communication channels so multiple models can work on shared repositories without collision. Rysh solves this by introducing a session-wide shared board interface. Panes execute models in isolation, but individual agents can post status updates, issue work orders, and read project milestones via the board without stealing terminal focus from the user. Commands like board post and board tail allow agents to publish updates tagged by pane identifier, enabling coordinated automation pipelines directly inside the terminal multiplexer.

This architecture enables complex graph engineering where agent relationships operate as directed workflows. In multi-agent fleet setups, a roadmap agent running Claude can post primary goals to the board, a fleet-manager running Codex can parse those goals into discrete task orders, and parallel worker panes can complete implementation tasks simultaneously. Workers can operate independently or receive direct corrections from manager agents before finalizing commits, allowing full software components like client-side applications to be constructed without manual human intervention for every intermediate step.

Rysh – an AI harness where Claude and Codex: who is affected

Engineers working on macOS and Linux environments receive native PTY support out of the box. For Windows developers, Rysh natively compiles to a windows/amd64 binary tagged as cli-only, but native Windows execution lacks a ConPTY implementation, leaving the PTY layer disabled. As a result, native Windows builds can run headless control commands like rysh send, rysh exec, and rysh doctor, but cannot open interactive session panes natively. Windows users requiring full terminal pane functionality must run Rysh inside WSL2 under Linux.

The project distribution maintains a clear distinction between the open-source binary rysh and the separate prebuilt distribution binary ry. The two names are intentionally non-interchangeable so both binaries can coexist on a single system without collision. Documentation and scripts written for rysh do not apply to ry, requiring users to verify their installed version using version check flags prior to executing automated maintenance scripts.

Rysh – an AI harness where Claude and Codex: what to watch

Initial setup requires initializing provider keys via the onboard command, which validates environment variables like ANTHROPIC_API_KEY, generates a project-local configuration file named rysh.config.yaml, and attaches the active session. Future developments will focus on expanding native Windows PTY capabilities and refining the shared board messaging protocols between autonomous agent panes. Developers can inspect system health and configuration status at any point by executing the rysh doctor utility from the command line.

As agentic CLI tools evolve, Rysh provides an Apache-2.0 open-source foundation for managing local multi-agent fleets within standard terminal workflows. Developers can monitor ongoing releases, inspect source code modifications, or submit contributions through the official GitHub repository workspace.