Your team adopted coding agents. Claude Code writes the feature. Codex reviews the diff. A script runs the tests. The problem: none of that happens the same way twice, there is no audit trail, and when something goes wrong, nobody can say what the agent actually did. Tencent just open-sourced LoopForge, a self-hosted platform that turns ad-hoc agent chats into repeatable delivery workflows with human approval gates.
What LoopForge Is #
LoopForge is an orchestration layer for coding agents β not a coding agent itself. It sits above tools like Claude Code, Codex, CodeBuddy, OpenCode, and Cursor CLI, and coordinates them through structured delivery workflows called SOPs (Standard Operating Procedures). The idea is simple: when you find a way of shipping that works, you codify it. Next time, the same process runs β with the same agents, the same scripts, and the same human checkpoints β without anyone having to remember the steps.
The project hit 1,243 GitHub stars and trended on November 15, gaining 484 stars in a single day. It is MIT licensed and currently at v0.3.6.
The SOP Engine #
Everything in LoopForge revolves around SOPs. A SOP is a graph of four node types:
- Agent nodes β run a coding agent (Claude Code, Codex, Kimi Code, DeepSeek Harness, or any of the ten supported CLIs) with a specific task
- Script nodes β deterministic shell steps: build, test, lint, generate a diff report
- Human gate nodes β the run; a human reviews the artifacts and approves, or sends the work back to an earlier step
- Sub-SOP nodes β embed another SOP as a nested workflow
A real feature delivery SOP might look like this: clarify requirements (human gate) β implement (Claude Code) β run tests (script) β independent review (Codex) β human sign-off (human gate) β ship (script). Every run is logged, checkpoint-resumable, and sandboxed in Docker.
Getting Started #
Setup requires Docker with Compose and Python 3. The installer handles Docker if it is missing.
git clone https://github.com/Tencent/LoopForge.git
cd LoopForge
bash scripts/install.sh
Open http://localhost:8090. The first account you create becomes the administrator. LoopForge ships with generic built-in SOPs that run without a repository, so you can test the workflow before wiring it to a real project.
One caveat: SOP runs use Docker sandbox containers through the host socket. That gives each run isolation but also means host daemon access. LoopForge is built for trusted internal teams, not public multi-tenant deployments.
The Self-Hosted Case #
Cloud-based coding agent platforms route prompts, code context, and agent outputs through third-party infrastructure. For many teams that is fine. For teams in finance, healthcare, or defense β or any company that takes IP protection seriously β it is not.
LoopForge keeps everything on your infrastructure. Every prompt, every agent response, every human gate decision stays inside your audit boundary. As enterprise coding agent adoption matures, that audit trail is going to matter more. AWS added native agent observability to Bedrock AgentCore this past summer for the same reason.
Why This Matters #
Most coding agent tools are still built for a single developer. LoopForge is the first serious answer to a question teams are starting to ask: how do we ship with agents at team scale, consistently and accountably?
The industry is moving through a predictable sequence. Individual developers adopted agents in 2024. Teams started using them in 2025 and ran into the chaos of no shared process. The governance layer is arriving in 2026 β LoopForge, tools like MakerChecker, cloud observability additions. This is what the maturation of the agent era looks like.
If your team is using coding agents ad-hoc today, LoopForge is worth a look. And if you are not thinking about process and audit trails for agent-driven development, you probably should be.
The repository is at github.com/Tencent/LoopForge, MIT licensed, with active development and ARM64 support in recent releases.