Show HN: AI Harness that lets a graph of Codex and Claude intercommunicate
Rysh CLI is an agentic terminal multiplexer written in Go that combines familiar terminal navigation (tabs, panes, splits, vim/htop) with AI agent capabilities in every pane Each pane can independently run AI agents (Claude, Codex) that answer prompts and call tools, with sessions persisting as daemons across restarts Supports multi-agent parallel workflows with cross-vendor agent orchestration (e.g., Claude and Codex working side-by-side in stacked panes) Includes SecretNAT for security: secret
Analysis
TL;DR
- Rysh CLI is an agentic terminal multiplexer written in Go that combines familiar terminal navigation (tabs, panes, splits, vim/htop) with AI agent capabilities in every pane
- Each pane can independently run AI agents (Claude, Codex) that answer prompts and call tools, with sessions persisting as daemons across restarts
- Supports multi-agent parallel workflows with cross-vendor agent orchestration (e.g., Claude and Codex working side-by-side in stacked panes)
- Includes SecretNAT for security: secrets are tokenized before leaving the machine, with reversible local mapping so models never see plaintext credentials
- Agent fleets can be designed as graphs with directed edges representing communication paths, enabling hierarchical task delegation and parallel execution
Why It Matters
This represents a significant evolution in how developers interact with AI agents—moving from single-agent chat interfaces to multi-agent terminal environments where agents can collaborate, delegate, and work in parallel. For AI practitioners, it demonstrates a practical architecture for orchestrating multiple LLM calls with persistent state, secure credential handling, and graph-based workflow design.
Technical Details
- Architecture: Go-based terminal multiplexer with daemon-persisted sessions; agents survive session restarts and maintain context (demonstrated via codeword recall)
- Multi-agent orchestration: Supports running multiple agents from different vendors (Claude, Codex) simultaneously in stacked panes, with
Ctrl+Snavigation and parallel file generation - Graph-based fleet design: Agent organizations are modeled as directed graphs where edges define communication paths; orders travel down as messages, results bubble up on a shared board
- SecretNAT security: On-by-default secret substitution replaces credentials with tokens in outbound requests; responses with plaintext credentials are reported to panes but not rewritten; local reversible mapping via
##snat get <token> - Dual binary distribution:
rysh(open-source, Apache-2.0) andry(prebuilt commercial distribution) are separate; requires Go 1.25.3+ for source builds; WSL2 supported on Windows
Industry Insight
- The daemon-based session model (vs. ephemeral terminal sessions) enables persistent agent state, which is critical for long-running multi-agent workflows that need to survive interruptions—this could become a standard pattern for production agent orchestration
- The graph-based agent topology (with explicit edges defining inter-agent communication) provides a programmable abstraction for complex multi-agent systems, moving beyond flat prompt chains to structured organizational designs
- SecretNAT's approach of tokenizing secrets at the client side before they reach LLM APIs addresses a growing security concern in AI adoption; this client-side secret management pattern could influence how future AI tools handle credential security by default
Disclaimer: The above content is generated by AI and is for reference only.