# LoopForge

> A self-hosted workspace that runs coding agents through repeatable delivery workflows with human approval gates

- Category: AI Coding / AI SDLC Automation
- Language: Go
- Licence: MIT
- Repository: https://github.com/Tencent/LoopForge
- Website: https://loopforge.cn
- Canonical page: https://ai-tldr.dev/tools/loopforge/
- Updated: 2026-11-12

## In short

LoopForge is an open-source, self-hosted workspace from Tencent for shipping software with coding agents. It sits in the AI SDLC Automation category of the AI Coding tools and is open-source, written in Go and licensed MIT. To get started: clone and start the platform.

## Overview

LoopForge is an open-source, self-hosted workspace from Tencent for shipping software with coding agents. Instead of driving one agent CLI by hand, you describe a task to Main Bot in a chat, and it works in a Docker sandbox with your repositories, hands parts of the job to project roles, and can open a pull request when it is done.

Once a way of working proves itself, LoopForge lets you save it as an SOP: a fixed workflow that mixes agent steps, deterministic scripts and human approval gates. A typical SOP clarifies a requirement, designs, implements, has an independent agent review the change, runs tests as a script, and then waits for a person to accept or send the work back with feedback. A visual designer builds these flows, and runs can be paused, resumed or taken over.

LoopForge orchestrates the coding-agent CLIs a team already uses, including CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI, Antigravity, Kimi Code, Pi, DeepSeek Harness and ZCode. A separate LoopForge Local mode installs the built-in workflows straight into a coding assistant such as Claude Code, Codex, CodeBuddy or Cursor, without running the platform.

## What it does

- Chat workbench where Main Bot works in a Docker sandbox with your repositories, with per-conversation executor, model and permission settings
- SOP workflows that combine agent roles, deterministic scripts, sub-SOPs, previews and human approval gates
- Visual SOP designer, plus pause, resume and take-over controls for running workflows
- Reusable roles with their own prompts, executors, rules, Skills and MCP tools
- Requirement tracking through to pull/merge request status, with live run logs, artifacts and a "Needs review" inbox
- Scheduled AI tasks and token and cost accounting
- Projects, member accounts and an OpenAPI-described HTTP and MCP API

## Getting started

The platform runs from Docker with one script; it needs Git, Docker with Compose, Python 3 and curl, and the installer can add missing pieces on Linux and macOS. API keys are only needed for chat and agent steps — the demo SOP runs without one.

1. Clone and start the platform

   Clone the repository and run the start script. It pulls the pinned platform image and brings the workspace up.

```bash
git clone https://github.com/Tencent/LoopForge.git
cd LoopForge
./start.sh
```

2. Manage the running instance

   The same script handles the lifecycle: `./start.sh logs`, `stop`, `restart` and `pull`, or `dev` to build from source. If downloads from ghcr.io stall, point it at a mirror.

```bash
MIRROR=ghcr.nju.edu.cn ./start.sh
```

3. Or use LoopForge Local inside your coding assistant

   To run the built-in workflows without the platform, install LoopForge Local into a project (Node.js 20+ and Python 3.8+ required), then start a workflow from Claude Code with `/start-devflow "<your request>"`.

```bash
npx -y loopforge-cli@latest install claude
```

## When to use it

- Turning a team's proven delivery process into a repeatable SOP that coding agents follow step by step
- Putting a human sign-off gate between an agent's finished change and the merge
- Running Claude Code, Codex or CodeBuddy in sandboxes on your own servers instead of a vendor cloud
- Tracking requirements through to pull request status and seeing which agent runs need review

## Version history

- 2026-11-12 v0.3.8: Renders sandbox dev servers in the side browser, asks before handing a requirement to an SOP, mirrors the platform image to Tencent's registry and stops trusting forwarded identity headers.
