orbit-cli 0.1.0

Local-first, multi-model DevOps CLI agent

Orbit

Local-first, multi-model DevOps CLI agent.

Orbit turns natural language into safe, observable DevOps actions. It decomposes goals into plans, picks the best local model for each step, and executes with a 4-tier safety system — all running against your own Ollama instance.

Features

• Multi-model routing — automatically selects the best locally-available model for each subtask (coding, ops, analysis)
• 4-tier safety system — every command is classified as safe, caution, destructive, or nuclear before execution
• Auto-replan — observes command output and adjusts the plan when things go wrong
• Context-aware — scans your git state, Docker containers, Kubernetes clusters, and system environment
• Command history — searchable SQLite-backed history of every action taken
• Runbook capture — record successful workflows as replayable YAML runbooks
• Streaming output — real-time LLM and subprocess output in a Rich terminal UI

Installation

pip install orbit-cli

Requirements

• Python 3.11+
• Ollama running locally with at least one model pulled (e.g. ollama pull qwen2.5:7b)

Quick Start

Execute a goal

orbit do "find large files in this repo and show disk usage by directory"

Orbit will decompose the goal, build a plan, ask for confirmation on risky steps, execute, and replan if needed.

Scan your environment

orbit sense

Collects context from git, Docker, Kubernetes, and system — shows what Orbit sees before planning.

Debug the last failure

orbit wtf

Analyzes the most recent failed command with full context and suggests fixes.

Ask a question

orbit ask "what Kubernetes pods are in CrashLoopBackOff and why?"

One-shot question answering with environment context.

Check configuration

orbit config doctor

Verifies Ollama connectivity, available models, and configuration health.

Safety Tiers

Tier	Behavior	Examples
safe	Execute silently	ls, cat, kubectl get, docker ps
caution	Single confirmation	git push, docker build, kubectl apply
destructive	Impact analysis + double confirm	rm, kubectl delete, git reset --hard
nuclear	Type confirmation + cooldown	Destructive in production context

Production detection automatically escalates destructive commands to nuclear when it detects a production environment (main/master branch, prod namespace, etc.).

Configuration

Orbit stores its configuration at ~/.orbit/config.toml. Run orbit config doctor to check your setup.

Key settings:

[llm]
provider = "ollama"          # ollama, openai, or anthropic
base_url = "http://localhost:11434"

[safety]
default_tier = "caution"     # for unrecognized commands

[agent]
max_steps = 15               # hard budget per goal
max_replans = 3              # retries per step
max_llm_calls = 25           # total LLM calls per goal

Optional Extras

pip install orbit-cli[rag]        # ChromaDB for semantic memory
pip install orbit-cli[openai]     # OpenAI provider
pip install orbit-cli[anthropic]  # Anthropic provider
pip install orbit-cli[all]        # everything

Development

git clone https://github.com/abhimanyubhagwati/orbit-cli.git
cd orbit-cli
pip install -e ".[dev]"
pytest
ruff check orbit/ tests/
mypy orbit/

License

Apache 2.0 — see LICENSE for details.