openBatarangs

Agentic coding CLI for local models.

Auto-discovers the best model on your hardware, thinks, reads, edits, and verifies — no manual model ops, no cloud required.

A DeepCode-style autonomous coding agent that runs on Jetson-class edge devices and anywhere Ollama runs.

---

openBatarangs is an agentic coding CLI for local LLMs. It talks to an Ollama server, auto-discovers the best installed model for your hardware, and iterates with tools to explore, edit, and verify code — no manual model selection required.

It is designed to run well on Jetson-class edge devices (Orin Nano/NX Super, 16 GB unified memory) and on any desktop/laptop that can run Ollama.

Why Ollama?

• Same llama.cpp + CUDA/ROCm/Metal engine used by most local LLM tools.
• Auto-discovers installed models through the Ollama API.
• No heavy Rust/CUDA compile per machine.
• Best performance-per-watt-per-dollar for local agentic coding on edge devices.

Performance per watt per dollar

openBatarangs is built around the local-first economics of edge AI:

• No cloud API fees. Every prompt stays on-device; the only cost is electricity.
• Unified memory wins on Jetson. The Orin 16 GB's GPU/CPU share memory, so a 4-7 GB Q4_K_M coding model fits without duplicating weights across VRAM and RAM.
• Small quantized models beat big cloud round-trips for agentic loops. A 3B-8B Q4 coding model is fast enough to iterate (read → edit → build → fix) while leaving headroom for a 32K+ context window.
• The auto-picker optimizes for watts, not just raw tokens/sec: it refuses models that don't fit comfortably, avoiding swap thrash and thermal throttling.
• Same binary runs on Jetson, laptops, and desktops — you only pay for the hardware you already own.

Exact tok/s depends on clocks, thermals, and model size; run openbatrangs doctor to see which model your hardware scores best for this workload.

Requirements

• Ollama installed (the CLI auto-starts ollama serve if it is installed but not running).
• If Ollama is not installed at all, run openbatrangs setup once — it installs Ollama and pulls a coding model for you.
• If no suitable model is installed, openBatarangs auto-pulls a recommended coding model unless you pass --no-auto-pull.

Install / Build

git clone https://github.com/arpanpathak/openbatrangs.git
cd openbatrangs

# Build release
cargo build --release

# Or install the binary into ~/.cargo/bin
cargo install --path .

The binary is at target/release/openbatrangs (or ~/.cargo/bin/openbatrangs).

Prebuilt binaries (easiest)

Download from the latest GitHub release:

# Linux x86_64 (most desktops/laptops)
curl -fsSL -o openbatrangs https://github.com/arpanpathak/openbatrangs/releases/latest/download/openbatrangs-x86_64-unknown-linux-gnu

# Linux aarch64 (Jetson, Raspberry Pi 5, Apple silicon Linux)
curl -fsSL -o openbatrangs https://github.com/arpanpathak/openbatrangs/releases/latest/download/openbatrangs-aarch64-unknown-linux-gnu

chmod +x openbatrangs
./openbatrangs setup
./openbatrangs

Or use the installer:

curl -fsSL https://github.com/arpanpathak/openbatrangs/releases/latest/download/install.sh | sh

Quick start

# One-time auto setup: install/start Ollama + pull a coding model
openbatrangs setup

# Interactive mode starts in chat (no tools); type /mode agent for full agent
openbatrangs

# One-shot agent mode with a task
openbatrangs "fix the Rust build errors"

# One-shot agent in a specific directory
openbatrangs --cwd /path/to/project "add a --dry-run flag"

# See what models are installed and which is best
openbatrangs list-models

# Check Ollama and get a recommendation
openbatrangs doctor

# Force a specific model
openbatrangs --model qwen2.5-coder:7b "explain this repo"

# Read-only mode (no file writes or shell commands)
openbatrangs --read-only "suggest a refactor plan"

# Ask before every write/command
openbatrangs --confirm "update the CLI docs"

Interactive REPL commands

Inside the openBatarangs> prompt:

/help          show all commands
/exit, /quit   leave the REPL
/setup         install/start Ollama + pull a model
/models        list installed models + scores
/model <tag>   switch model (e.g. /model qwen2.5-coder:7b)
/read-only     toggle read-only mode
/confirm       toggle confirm-before-write/command
/steps <n>     set max agent steps
/cwd <path>    change workspace
/doctor        check Ollama + best model
/mouse on|off  wheel/scrollbar capture (default on)

The interactive TUI starts in chat mode (no tools). Use /mode agent to enable the full agentic loop, or /mode plan for read-only planning.

Anything else you type is sent to chat; in agent/plan mode it is sent to the coding agent as a task.

Agent tools

The agent can use these tools during a task. It does not scan the workspace unless your task explicitly asks about the current directory, project structure, codebase, or which files exist:

Tool	Description
list_files	List files in a directory (max depth 2; skips target, .git, node_modules, data, build/cache dirs, etc.)
read_file	Read a text file with a size cap
grep_files	Regex search across workspace files
write_file	Write or overwrite a file (relative paths only)
run_command	Run a shell command in the workspace (e.g. cargo check)
finish	Signal the task is complete

For safety:

• Tool paths must be relative; absolute paths and .. are rejected.
• --read-only disables write_file and run_command.
• --confirm asks before each write/command.

Model auto-discovery

openBatarangs scores installed models by:

• Memory fit (model file size vs. system memory)
• Parameter size sweet spot for agentic coding (roughly 3B–8B on 16 GB devices)
• Coding-model name bonus (qwen2.5-coder, deepseek-coder, etc.)
• Context window (target ~32K)
• Quantization quality

If nothing suitable is installed, it can automatically pull a recommended model (qwen2.5-coder:7b on >=12 GB systems, otherwise qwen2.5-coder:3b).

Common flags

--ollama-url <URL>   Ollama server URL (default http://localhost:11434)
--model <TAG>        Use a specific Ollama model tag
--cwd <DIR>          Workspace directory (default .)
--max-steps <N>      Max agent iterations (default 12)
--min-context <N>    Minimum context window for auto-selection (default 8192)
--read-only          Disable writes and shell commands
--confirm            Ask before writes/commands
--no-auto-pull       Never auto-pull models

Roadmap

This is the first working version. Future steps for standalone distribution:

• Prebuilt binaries for aarch64 and x86_64
• cargo install from crates.io
• Support OpenAI-compatible remote endpoints in addition to Ollama
• Optional direct GGUF fallback via mistralrs or llama.cpp
• Better token budgeting / context compression for long agent sessions
• Installer script that checks for Ollama and installs a recommended model

License

AGPL-3.0. Contributions are welcome; see CONTRIBUTING.md.