Self-Hosted Docker Sandboxes Where Coding Agents Actually Ship Code
You've probably tried pointing a coding agent at a repo and watching it produce something that looks like a PR. Then you spend the next hour figuring out what it actually changed, whether it broke anything, and how to hand it off for review. The agent did the typing; you did all the trust-building. LoopForge is a self-hosted workspace from Tencent that tries to close that gap—running coding agents inside Docker sandboxes, wiring them into defined roles, and putting human approval gates in front of whatever ships.
What It Does
LoopForge is a self-hosted platform where your team delivers software with coding agents. You describe what you need to a chat interface called Main Bot, and it spins up a Docker sandbox, checks out your repositories, delegates work to project roles, and opens the PR. The coding-agent CLIs do the actual work—CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI, Antigravity, Kimi Code, Pi, DeepSeek Harness, and ZCode are all supported. LoopForge layers the roles, sandboxes, checkpoints, and history around them, all on your own infrastructure under the MIT license.
The interesting part is the second mode. When a way of working proves itself, you can save it as an SOP—a standard workflow of agents, scripts, and human gates. The engine then runs Agent, Script, Human-gate, and sub-SOP nodes in order, each in its own sandbox. Both paths track the same requirements, so every change links back to what was originally asked for and counts toward delivery and cost reports.
Why It's Cool
It treats agent work as a process, not a prompt. Most agent tooling assumes one long conversation and a lot of hope. LoopForge separates the two modes cleanly: chat with Main Bot for exploring and quick fixes, or run an SOP when you need fixed stages and sign-off. That distinction maps to how teams actually work—some tasks are a conversation, some are a checklist.
The human gates are structural, not an afterthought. In SOP mode you approve at human gates, send work back to a specific step, retry, or rerun from any node. You can also pause, resume, or take over a running SOP from chat. That's a real control surface, not a "please review this diff" message at the end.
Sandboxes per workspace. Main Bot opens a conversation workspace—a Docker sandbox with your repos—on demand and calls project roles as subagents. SOP nodes get their own sandbox too. You're not letting an agent loose on your host machine and hoping for the best.
Bring your own agent. The list of supported CLIs is long and covers the major players. LoopForge isn't selling you a model; it's selling the scaffolding around whichever model you already trust.
Everything stays on your infrastructure. Self-hosted, Docker-based, MIT licensed. For teams that can't send code to a third-party service, that's not a nice-to-have.
How to Try It
The README points to a quick-start section and a full documentation set, so start there rather than guessing at commands:
- Head to the repository and read the quick-start section in the README.
- Check the documentation linked from the README for setup details, including conversation workspaces.
- There's also an 80-second product tour video embedded in the README if you want to see the flow before installing anything.
- Note that the README also mentions a "Local (Legacy)" mode—worth reading if you're evaluating whether the Docker-based setup is the right fit for you.
Since LoopForge is self-hosted with Docker, expect to have Docker available on whatever machine you're running it on. The README covers the specifics; I'd rather point you there than invent install steps that don't match reality.
Final Thoughts
LoopForge is aimed at teams that have already decided coding agents are useful but are tired of the manual overhead around them—the sandboxing, the review handoffs, the "how did we do this last time" problem. The SOP concept is the real differentiator here: it turns a one-off successful agent session into something you can run again with the same stages and the same checkpoints. If you're a solo developer doing quick fixes, the chat mode alone might be enough. If you're running a team where repeatability and approval matter, the SOP engine is the part worth evaluating. Either way, it's MIT-licensed and runs on your own hardware, which makes trying it a low-stakes proposition.
Follow @githubprojects for more developer tools and open source projects.