The software factory for small teams of expert engineers

Arc replaces the prompt-wait-review loop with version-controlled workflow graphs that orchestrate AI agents, shell commands, and human decisions into repeatable, long-horizon coding processes.

Define workflows in Graphviz DOT, route tasks to the right model with CSS-like stylesheets, and let the engine handle orchestration, parallelism, checkpointing, and verification -- all from a single Rust binary.

---

Table of Contents

• Key Features
• Quick Start
• Example Workflow
• Supported Models
• System Requirements
• Help or Feedback
• Contributing
• License

---

Key Features

What Arc Does

	Feature	Description
🤖	Multi-model orchestration	Route tasks to the right model per node -- cheap for boilerplate, frontier for hard reasoning
💠	Declarative workflows	Define pipelines in Graphviz DOT -- diffable, reviewable, composable
✋	Human-in-the-loop	Approval gates, interviews, and real-time steering let you intervene at the right moments
🔄	Checkpoint and resume	Git-native checkpointing after every stage -- inspect, revert, or fork from any point
🛡️	Adaptive verification	Combine LLM-as-judge, test suites, and human review into quality gates
📊	Full observability	Every tool call, agent turn, and decision point captured in a unified event stream

How Arc Does It

	Feature	Description
🎨	Model stylesheets	CSS-like selectors (*, .class, #id) assign models, providers, and reasoning effort
🌳	Git-native checkpoints	Every stage commits to a branch -- resume interrupted runs exactly where they left off
📦	Sandbox isolation	Run agent tools in local, Docker, Daytona cloud VMs, or exe.dev ephemeral VMs
🔌	MCP integration	Extend agents with any Model Context Protocol server (Playwright, databases, APIs)
🔁	Loops and fan-out	Implement-test-fix cycles, parallel code reviews, and ensemble multi-provider patterns
🦀	Written in Rust	Single compiled binary with minimal dependencies -- no Python runtime, no npm install
⚙️	CLI and API modes	arc run for local dev, arc serve for production with a React web UI

Read the full documentation for details.

---

Quick Start

Warning

Arc is in private research preview. Contact bryan@qlty.sh if you're interested in trying it.

Install

Download the latest release for your platform:

# macOS (Apple Silicon)
curl -fsSL https://github.com/brynary/arc/releases/latest/download/arc-aarch64-apple-darwin.tar.gz | tar xz
sudo mv arc /usr/local/bin/

# Linux (x86_64)
curl -fsSL https://github.com/brynary/arc/releases/latest/download/arc-x86_64-unknown-linux-gnu.tar.gz | tar xz
sudo mv arc /usr/local/bin/

Or download directly from GitHub Releases.

Build from source

Requires Rust (latest stable):

git clone https://github.com/brynary/arc.git
cd arc
cargo build --release
# Binary is at ./target/release/arc

Run the setup wizard

arc install

Initialize your project

cd my-repo/
arc init

Run your first workflow

arc run hello

---

Example Workflow

A plan-approve-implement workflow where a human reviews the plan before the agent writes code:

digraph PlanImplement {
    graph [goal="Plan, approve, implement, and simplify a change"]

    start [shape=Mdiamond, label="Start"]
    exit  [shape=Msquare, label="Exit"]

    plan      [label="Plan", prompt="Analyze the goal and codebase. Write a step-by-step plan.", reasoning_effort="high"]
    approve   [shape=hexagon, label="Approve Plan"]
    implement [label="Implement", prompt="Read plan.md and implement every step."]
    simplify  [label="Simplify", prompt="Review the changes for clarity and correctness."]

    start -> plan -> approve
    approve -> implement [label="[A] Approve"]
    approve -> plan      [label="[R] Revise"]
    implement -> simplify -> exit
}

Node types at a glance

Shape	Type	What it does
Mdiamond	Start	Workflow entry point
Msquare	Exit	Workflow terminal
box (default)	Agent	Multi-turn LLM with tool access
tab	Prompt	Single LLM call, no tools
parallelogram	Command	Runs a shell script
hexagon	Human gate	Pauses for human input
diamond	Conditional	Routes based on conditions
component	Parallel	Fans out to concurrent branches
tripleoctagon	Merge	Collects parallel branch results

Multi-model routing with stylesheets

graph [
    model_stylesheet="
        *        { llm_model: claude-haiku-4-5; reasoning_effort: low; }
        .coding  { llm_model: claude-sonnet-4-5; reasoning_effort: high; }
        #review  { llm_model: gemini-3.1-pro-preview; llm_provider: gemini; }
    "
]

Selectors follow CSS specificity: * (0) < shape (1) < .class (2) < #id (3).

---

Supported Models

Arc supports models from Anthropic, OpenAI, Gemini, Kimi, Zai, MiniMax, and Inception. Route tasks to the right model per node and configure provider fallback chains per-run.

Run arc model list locally or see the Models documentation for the full catalog.

---

System Requirements

• macOS (Apple Silicon) or Linux (x86_64)
• At least one LLM provider API key (Anthropic, OpenAI, or Gemini)
• Git (for checkpoint and resume)
• Rust (only if building from source)

---

Help or Feedback

• Read the documentation
• Bug reports via GitHub Issues
• Feature requests via GitHub Issues
• Email bryan@qlty.sh for access or questions

---

Contributing

See CLAUDE.md for build commands, architecture overview, and development conventions.

---

License

Arc is licensed under the MIT License.