RARA

A fast, local-first coding agent for your terminal.

RARA runs directly on your machine — no cloud dependency required. Pick your model, ask it to write code or explore your repo, and watch it work through a rich TUI with live streaming output, syntax highlighting, and side-by-side context.

---

Why RARA

• Runs locally. Your code, your models, your disk. Nothing leaves your machine unless you choose a hosted provider.
• Model freedom. Use Claude, DeepSeek, Gemini, Ollama, any OpenAI-compatible endpoint, or local Candle models — switch providers and models mid-session with /model.
• Full tool suite. Edit files, run shell commands, search code, browse the web, spawn sub-agents, update project memory, and load local skills.
• Sandbox by default. On macOS and Linux, command execution is wrapped in a platform sandbox; you control network and filesystem access per command.
• Session persistence. Every thread is saved, restorable, forkable, and searchable — pick up where you left off.

---

Quick Start

git clone https://github.com/linkerdog/rara.git
cd rara
cargo run -- tui

Choose a provider on first launch — OpenAI, DeepSeek, Ollama, or a local Candle model with zero setup.

For RARA_API_KEY, set the environment variable or pass --api-key.

# Start a TUI session with a specific provider
rara --provider deepseek --model deepseek-flash tui

# Ask a one-shot question
rara ask "summarize this repo"

# Resume your last session
rara resume --last

# Start this session with local permission checks bypassed
rara --dangerously-skip-permissions

---

TUI

The terminal UI gives you a full development environment:

• Streaming responses — watch the agent think and act in real time.
• Live bash output — stdout/stderr appear as the command runs.
• Syntax highlighting — code blocks, diffs, and inline code are colorized.
• Sidebar context — current-turn token usage, model info, open files, and child sessions at a glance.
• Slash commands — /model to switch providers, /status for runtime state, /context for prompt diagnostics, /help for available commands. Inspection commands remain available during work; runtime mutations show a disabled reason until the current task finishes.
• Permission controls — /permissions shows the effective policy and pending changes. Full access (always allow) skips local tool approval and classifier checks and enables sandbox network access. Start directly in this mode with --dangerously-skip-permissions; the flag does not change saved defaults or disable OS/container isolation.
• Approval card — when a shell command needs approval, use Up/Down plus Enter, or press 1-4, directly on the transcript card.
• Follow-up queuing — type ahead while the agent is busy; your messages queue and process in order.

---

Providers

Provider state includes API key, base URL, model name, reasoning effort, and context window size. OpenAI-compatible providers can be saved as named profiles for quick switching.

Additional compatible services can be configured using OpenCode-style provider and model maps in rara.json or rara.jsonc. Presets include Groq, Together, xAI, Mistral, MiniMax, Z.ai, and Hyperbolic; custom API roots are also supported.

{
  "provider": {
    "groq": {
      "models": {
        "llama-3.3-70b-versatile": { "name": "Llama 3.3 70B" }
      }
    }
  }
}

Use /connect to save the provider key, then /model to select a model, or set GROQ_API_KEY and pass --model groq/llama-3.3-70b-versatile. Global provider documents live in $RARA_HOME (default ~/.rara); project documents override global fields. See the configuration contract for merge precedence, variable references, and the supported compatibility scope.

---

Tools

RARA gives the agent access to your workspace:

A sidechannel LLM classifier (using a lightweight auxiliary model) evaluates shell and web tool calls before execution — blocking dangerous commands and auto-approving safe ones, following the same pattern as Claude Code's YOLO classifier.

Long tool output is folded in the transcript with a preview and a path to the full result.

---

Sandbox

Command execution is sandboxed by default:

PTY sessions on macOS currently run outside the seatbelt sandbox due to PTY/seatbelt interaction constraints.

---

Where State Lives

• $RARA_HOME (default ~/.rara/) — configuration, profiles, and credentials
• ~/.cache/rara/ — model downloads, runtime metadata
• ~/.local/share/rara/ — workspace thread logs and session data
• <workspace>/.rara/ — per-project memory and lock files

---

License

Apache 2.0