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
• Architecture
• 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.

Prerequisites

• Rust (latest stable)
• At least one LLM API key (ANTHROPIC_API_KEY, OPENAI_API_KEY, or GEMINI_API_KEY)

Clone and build

git clone https://github.com/qltysh/arc.git
cd arc
cargo build --release

Configure API keys

cp .env.example .env
# Edit .env and add at least one provider key

Run the setup wizard

./target/release/arc install

Verify your installation

./target/release/arc doctor --live

Run your first workflow

./target/release/arc run docs-internal/demo/01-hello.dot

Or try a multi-step workflow with a human approval gate:

./target/release/arc run docs-internal/demo/10-plan-implement.dot

---

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

Model	Provider	Aliases
claude-opus-4-6	Anthropic	opus
claude-sonnet-4-5	Anthropic	sonnet
claude-haiku-4-5	Anthropic	haiku
gpt-5.2	OpenAI	gpt5
gpt-5.3-codex	OpenAI	codex
gpt-5.4	OpenAI	gpt54
gemini-3.1-pro-preview	Gemini	gemini-pro
gemini-3-flash-preview	Gemini	gemini-flash
kimi-k2.5	Kimi	kimi
glm-4.7	Zai	glm
minimax-m2.5	MiniMax	minimax
mercury-2	Inception	mercury

Run arc model list for the full catalog. Provider fallback chains are configurable per-run.

---

Architecture

Arc provides two interfaces backed by the same workflow engine:

Feature	CLI mode (arc run)	API mode (arc serve)
Execution	Synchronous, single run	Async, queued with scheduler
Human-in-the-loop	Terminal prompts	HTTP endpoints + web UI
Events	Printed to stderr	SSE stream
Concurrency	One run per process	Configurable (default 5)
Best for	Development, one-off runs	Production, integrations

Rust crates

Crate	Purpose
arc-cli	CLI entry point (run, exec, serve, doctor, install)
arc-workflows	Core workflow engine -- DOT parsing, stage execution, checkpoints
arc-agent	AI coding agent with tool use (Bash, Read, Write, Edit, Grep)
arc-api	Axum HTTP server with SSE event streaming
arc-llm	Unified LLM client -- Anthropic, OpenAI, Gemini, and more
arc-types	Auto-generated types from OpenAPI spec
arc-github	GitHub App auth, PR creation, checkpoint pushing
arc-db	SQLite with WAL mode and schema migrations
arc-mcp	Model Context Protocol client/server

TypeScript

Package	Purpose
apps/arc-web	React 19 + React Router + Vite + Tailwind frontend
arc-api-client	Auto-generated TypeScript client from OpenAPI spec

---

System Requirements

• macOS or Linux (x86_64 or ARM)
• Rust (latest stable) for building from source
• At least one LLM provider API key (Anthropic, OpenAI, or Gemini)
• Git (for checkpoint and resume)
• Docker (optional, for Docker sandbox mode)

---

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

Developing the CLI

Arc requires a working Rust toolchain:

git clone https://github.com/qltysh/arc.git
cd arc
cargo build --workspace
cargo test --workspace

Developing the web UI

cd apps/arc-web
bun install
bun run dev

Useful commands

Command	Description
cargo build --workspace	Build all crates
cargo test --workspace	Run all tests
cargo test -p arc-api	Test a single crate
cargo fmt --check --all	Check formatting
cargo clippy --workspace -- -D warnings	Lint
cd apps/arc-web && bun run typecheck	Type check the frontend

---

License

Arc is licensed under the MIT License.