Tencent/LoopForge

What it does

LoopForge is a self-hosted workspace for delivering software with coding agents. A Main Bot receives a request in chat, opens a Docker sandbox with your repositories checked out, delegates work to project roles, and opens a PR or merges into a shared branch. Repeatable work can be captured as an SOP — a workflow of agent, script, human-gate and sub-SOP nodes — so every requirement follows the same stages and sign-off.

The project does not replace CodeBuddy, Codex or Claude Code; it orchestrates them. Each agent step runs one of those CLIs, while LoopForge supplies the roles (prompt, model, rules, Skills, MCP tools), sandboxes, checkpoints, run history and approval inbox around them. It targets teams that want consistent, auditable agent-assisted delivery rather than one-off agent sessions.

Quick start

Prepare Git, clone the repository, then run the installer:

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

./start.sh pulls the GHCR release image pinned in compose.yaml instead of building locally. It checks Docker, Compose, Python 3 and curl, creating MySQL and the platform services, generating secrets, and installing built-in sandboxes, starter SOPs and a welcome project. Set LOOPFORGE_YES=1 to accept prompts non-interactively; macOS needs Docker Desktop and a working Python 3.

Open the address printed by the installer and create the first account, which becomes administrator. The service listens on 0.0.0.0:8090 by default. To try it without model credentials, switch to the demo project, open Platform check · script & approval, click Run SOP, and approve the waiting gate from Needs review. For real work, add an executor API key in personal settings, and connect GitHub if you want PRs against existing repositories.

Alternatively, install the CLI into an existing project to use the built-in workflows inside Claude Code or other supported assistants:

npx -y loopforge-cli@latest install claude

Then run /start-devflow "<your request>" in Claude Code.

Constraints

• The default Compose setup mounts the host Docker socket, so sandbox containers can control the host daemon; run it for trusted teams, not as untrusted multi-tenant isolation. See the deployment boundaries for the separate-worker option.
• New installations listen on all interfaces and assume a trusted team: signed-in users can read project data across the instance, and signup is open unless AGENTFLOW_ALLOW_SIGNUP=false is set. Restrict access and add HTTPS before exposing it beyond the team.
• Model usage is bring-your-own-key and billed externally; the repository only tracks token costs at public list prices. CodeBuddy and Codex ship in the Docker image, while Claude Code installs from npm on first use.
• The platform requires Docker and Python 3; LoopForge Local requires Node.js 20+, Python 3.8+ and a supported coding assistant. Local currently runs the built-in workflows only — exporting a platform SOP for local execution is planned.
• Management pages are bilingual, but the chat workbench is currently Chinese only. The repository is MIT-licensed; third-party components keep their own notices.

HotRepo's take

On 2026-11-11 LoopForge appeared on the GitHub Go daily chart at position 5 with 76 stars gained that day and 282 stars in total, so its recorded visibility is a single-day Go-language ranking rather than a sustained all-language trend. The rank is supported by an actual shipped implementation: a Go platform, a separate local CLI package, an end-to-end acceptance test, and a validation record with known gaps. What stands out for readers is the explicit human-gate model and Docker sandboxing as the delivery interface for coding agents — the same design that makes host Docker access and open signup the most important operational constraints to review before adopting it.

Project source and documentation