rho

rho is a fast, clean, local, private, and secure coding agent CLI built in Rust on Rig.

---

Installation

cargo install rho

---

Quick Start

# Start interactive REPL
rho

# Run one-shot prompt
rho -p "summarize this repository"

# Select model in <provider>/<model> format
rho --model anthropic/claude-3-7-sonnet
rho --model gemini/gemini-2.5-flash

# Resume a previous session
rho --resume <SESSION_ID>
# Or browse recent sessions interactively
rho --resume-picker

---

Built-in Tools

rho includes 10 fast, native tools designed for coding agents:

• read: Read file contents with line numbering, offset, and limit safeguards. Supports universal virtual file system schemes (diff://staged, diff://workdir, conflict://, pr://<number>), image sniffing, and downscaling.
• write: Create or overwrite files, automatically creating missing parent directories.
• edit: Apply targeted, exact text replacements to files.
• bash: Execute shell commands with process group cleanup, timeout protection, and binary sanitization.
• fd: Fast, gitignore-aware workspace file discovery with smart-case regex matching.
• rg: Fast, line-oriented content searching; gitignore-aware, skips binary files, and bounds output.
• subagent: Spawn isolated subagents with specialized roles (scout, critic, planner, general) to perform focused research, trade-off analysis, or code review without parent context bloat.
• lsp: Query Language Server Protocol servers for compiler diagnostics, symbol definitions, cross-file references, document symbols, and rename diff previews without invoking slow external shell build commands.
• web_search: Search the web and retrieve structured summaries and URLs. Supports selectable search engines (Brave, DuckDuckGo Lite, Yahoo, Firecrawl).
• web_fetch: Fetch and extract clean markdown, text, HTML, CSV, or feeds from web URLs. Includes specialized extractors for GitHub (issues, PRs with diffs, commits, raw code, directory trees) and YouTube (metadata and timed dialogue transcripts).

Web tools, search providers, MCP, and permissions can be configured in config.toml or interactively toggled via /settings under Tools & Permissions.

---

Feature Highlights & Documentation

Comprehensive guides are organized in docs/:

• Providers, Configuration & Skills

  • Authentication for 14+ providers (ChatGPT, Claude, Copilot, Antigravity, Anthropic, Gemini, DeepSeek, Local Ollama, and more).
  • Custom OpenAI-compatible endpoints with private network protection.
  • Hierarchical instruction loading (AGENTS.md, CLAUDE.md, .cursorrules).
  • Declarative parameter-guided skills (SKILL.md).

• Interactive UI & Terminal Styling

  • Dynamic upward-expanding multiline editor with stable screen scrollback.
  • Live two-line footer metrics: token count, context percentage, speed, cost, and active model.
  • In-memory FIFO message queueing (Alt+Enter) and non-blocking background turns.
  • Universal native theme that adopts your terminal's 16-color palette, with SGR-dim secondary text, terminal-detected card fills, or configurable outline borders.
  • Fenced Mermaid diagram rendering (Flowcharts with subgraphs, Sequence, State, Class, ER).

• Keyboard Shortcuts & Controls

  • Comprehensive keybinding reference organized by session flow, queueing, model controls, and editor navigation.
  • Custom keybinding configuration (~/.config/rho/keybindings.toml).

• Permissions, Privacy & Safety

  • Strict zero-telemetry policy: rho collects nothing, no analytics, no phone-home pings.
  • In-process safety layer separating baseline safe inspection from mutating commands.
  • Delegated Guard Model classification: automatically classifies shell commands using a fast local model (qwen2.5-coder:7b via Ollama) to execute safe commands without friction while intercepting hazardous operations.
  • Interactive approval modals: Allow, Edit (with multiline arrow navigation), Always (with pattern matching), and Deny (with feedback).
  • Fail-closed execution in headless automation and on guard model errors or timeouts.

• MCP Servers & Lifecycle Hooks

  • Full Model Context Protocol (MCP) support: stdio and streamable-http transports with OAuth 2.1 PKCE authorization.
  • Ecosystem interoperability with ~/.agents/mcp.json, project .agents/mcp.json, and .mcp.json.
  • Context budget tool gating (direct vs gateway vs auto), /mcp TUI modal, and rho mcp management suite.
  • One-shot lifecycle hooks in .agents/hooks/ for deterministic policy enforcement, tool gating, argument rewriting, and turn control without background daemons.

• Collab (Peer-to-Peer Live Pairing)

  • Share active sessions terminal-to-terminal over end-to-end encrypted QUIC powered by Iroh.
  • Capability-based cryptographic isolation: full interactive co-pilot vs. read-only spectator.
  • REPL slash commands: /collab start, /collab peers, /collab kick <id>, /collab rotate, /collab stop.
  • Collaborators join instantly with zero server setup via rho join <ticket>.
  • Full interactive TUI for guests: alternate-screen rendering, markdown token streaming, styled tool cards, and snapshot transcript hydration.
  • Shared interactive tool approval synchronization across terminals via in-TUI confirmation modals.

• Sovereign Agent Mesh (Peer-to-Peer Task Delegation)

  • Delegate autonomous subagent tasks across sovereign peer nodes over end-to-end encrypted Iroh QUIC streams.
  • Cryptographic capability tickets (rho-agent://) enforcing strict sandbox invariants: read_only (read/lsp/search only), propose (file diffs/edits without shell), or full access.
  • Start a worker node with rho agent serve --cap <read_only|propose|full>.
  • Delegate tasks from any terminal with rho agent run <ticket> "<prompt>".

---

Privacy & Zero Telemetry

rho is built from the ground up with a strict privacy-first architecture:

• Zero Telemetry: rho collects nothing. There are no analytics, no telemetry, no tracking pixels, no crash reporters, and no phone-home pings.
• Direct-to-Provider: Network requests travel strictly and directly between your machine and your configured model provider (e.g. Anthropic, OpenAI, Gemini, local Ollama). No prompts, source code, or completions are ever routed through intermediary proxies or secondary servers.
• 100% Local Storage: All authentication credentials, session histories, transcripts, and context caches reside exclusively on your local filesystem (in ~/.config/rho, ~/.local/share/rho, and project .rho/).
• Open Source & Auditable: The entire codebase is open source under MIT / Apache-2.0 and contains zero analytics SDKs or tracking dependencies.

---

Development

Run the test suite, linter, and format checks:

cargo test --workspace
make clippy
cargo fmt --all -- --check

---

Acknowledgments

• Special thanks to pi.dev for inspiring aspects of rho's design and ergonomics.
• Built on Rig for modular, performant LLM application pipelines in Rust.
• Powered by Iroh for direct, zero-infrastructure peer-to-peer QUIC networking and NAT traversal.