AgentFlow v0.2.0-beta.1 — See what's new
Built on Tauri + React + Rust

Visual AI Agent
Pipeline Builder

Build, orchestrate, and execute Claude Code agent pipelines with a drag-and-drop canvas. A desktop app for teams managing multi-step AI automation workflows.

Explore FeaturesView on GitHub
Start Pipeline
AI: Code Review
Shell: Run Tests
Approval Gate
Git: Push & Deploy

Features

Everything you need to orchestrate AI agents

From visual pipeline design to cost-aware execution, AgentFlow gives your team full control over complex Claude Code workflows.

Visual Pipeline Builder

Canvas-first editor with React Flow. Drag and drop 8 node types — AI Task, Shell, Git, Parallel, Loop, Approval Gate, Sub-pipeline, and Comment — connect them with conditional edges, and scaffold fast with 8 starter templates.

Drag-and-drop canvas8 node typesLoop iterationStarter templates

Execution & Monitoring

Live execution engine that spawns Claude Code CLI and shell processes with real-time stdout/stderr streaming. Track every run with SQLite-backed history, retry failures automatically, and resume from the exact point of failure.

Real-time log streamingRetry & resumeRun historyDesktop notifications

Cost & AI Control

Per-node cost tracking with budget limits that halt execution when exceeded. Choose the right Claude model for each node — Opus, Sonnet, or Haiku — and skip re-execution of unchanged nodes with output caching.

Budget limitsPer-node model selectionOutput cachingData passing

Safety & Governance

Approval gates pause execution for human review at critical steps. Secret variables stay masked in the UI and excluded from logs. Pre-run validation catches circular dependencies and missing references before execution starts.

Human-in-the-loopSecret variablesPipeline validationNode data passing

Agent Management

Browse, create, and edit Claude Code agents with a rich Markdown editor and live preview. Auto-discovers agents from your project's .claude/agents/ directory for immediate use in pipelines.

Markdown editorAuto-discoveryLive previewIn-pipeline use

Git-Native & Team-Ready

Pipelines are JSON, agents are Markdown — all stored in .claude/ for native version control. Developers can use the visual app or CLI interchangeably. Push pipelines via git for instant team sharing.

JSON + Markdown filesNo external serverCLI interopGit sharing

Node Types

Eight powerful building blocks

Combine these node types to build any workflow — from simple automation to complex multi-stage pipelines with loops, human oversight, and more.

AI Task

AI-powered code generation, review, and refactoring

$0.42
"Review this PR for security issues and suggest fixes"

Shell Command

Run tests, build projects, deploy, or any CLI operation

npm run test -- --coverage

Git Operation

Commit, push, branch, merge — automate your git workflow

git commit -m "feat: auto-generated changes"

Parallel

Run branches simultaneously, wait for all to complete

Run linting, tests, and type-check in parallel

Loop

Iterate over a list, running steps for each item

Loop over [file1, file2, file3] and review each

Approval Gate

Pause for human review at critical pipeline steps

"Review AI changes before merging to main"

Sub-pipeline

Execute another saved pipeline as a reusable module

Call "deploy-staging" pipeline as a step

Comment

Annotate pipeline sections with notes and documentation

Then
Success
Failure

How It Works

From project to production in four steps

Step 01

Point to Your Project

Open AgentFlow and select your git repository. It auto-detects existing Claude Code agents and pipelines in the .claude/ directory.

Step 02

Design Your Pipeline

Drag nodes onto the canvas, configure each step's instructions, and connect them with conditional edges for success/failure branching.

Step 03

Execute & Monitor

Hit Run and watch your pipeline execute in real-time. Stream live logs, track costs, and approve gates as they fire.

Step 04

Review & Iterate

Browse run history with full cost breakdowns. Resume failed runs, optimize your pipelines, and share via git push.

Architecture

How data flows through AgentFlow

A type-safe pipeline from React UI to Rust backend to Claude Code CLI, with real-time event streaming back to the canvas.

React UI

Canvas, Panels, Zustand Stores

Zustand store updates & user actions

Tauri IPC Bridge

invoke() commands & event listeners

Serialized command payloads

Rust Executor

Tokio async runtime, SQLx queries

Spawns child processes

Claude Code CLI + Shell

AI inference, bash execution, git ops

Emits events with logs & status

SQLite Database

Run history, cost tracking, settings

Project File Layout

your-project/

├── .claude/

│ ├── agents/← Agent markdown files

│ │ ├── code-reviewer.md

│ │ └── test-runner.md

│ └── pipelines/← Pipeline JSON files

│ └── review-and-deploy.json

├── src/

└── .mcp.json

Tech Stack

Modern, type-safe, and performant

A carefully chosen stack that prioritizes developer experience, compile-time guarantees, and cross-platform reliability.

Frontend

React 19UI library with concurrent features
TypeScript 5.9End-to-end type safety
Vite 7Lightning-fast dev server & builds
React FlowInteractive node-based canvas
ZustandLightweight state management
Tailwind CSS 4Utility-first styling

Backend

Tauri v2Lightweight desktop framework
RustMemory-safe, high-performance backend
TokioAsync runtime for parallel execution
SQLx + SQLiteType-safe database queries
SerdeSerialization / deserialization
NotifyFile system watching for live reload

Keyboard-First

Built for speed

Every action has a keyboard shortcut. Stay in flow without reaching for the mouse.

Ctrl+R
Run pipeline
Ctrl+S
Save pipeline
Ctrl+Z
Undo
Ctrl+Shift+Z
Redo
Space
Zoom to fit
Esc
Close panel
Ctrl+E
Export JSON
Del
Delete node
Ctrl+C
Copy node
Ctrl+V
Paste node

Guidance

When to use AgentFlow vs Claude CLI directly

Use AgentFlow

AgentFlow is designed for complex, multi-step workflows — pipelines with branching logic, parallel execution, approval gates, and retry policies. For these orchestration-heavy tasks, the visual builder, run history, and resume-from-failure capabilities save significant time compared to manually scripting the same flow.

  • Multi-step CI/CD pipelines
  • Workflows needing human approval
  • Parallel tasks with retry logic
  • Repeatable automations you want to track and share

Use Claude CLI directly

For simple, single-purpose tasks — quick bug fixes, one-off code generation, or exploratory questions — running Claude CLI directly is faster and more straightforward. Not every task needs a pipeline.

  • Quick bug fixes
  • One-off code generation
  • Exploratory questions
  • Simple single-step tasks

Ready to orchestrate your AI agents?

AgentFlow is open source and free. Clone the repo, build the app, and start automating your Claude Code workflows today.

Download for FreeRead the Docs