Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseCode / SDK · ActionableGitHub Trending AI→ original

Tencent Open-Sources LoopForge: Ten Coding Agents Run Under One Set of Rules on Your Own Server

A coding agent asked to "fix the bug" may skip the plan, forget the tests, or abandon the job halfway through, and nobody notices. Tencent has released LoopForge under the MIT license, a platform that runs 10 coding agents, including Codex and Claude Code, on your own server under one standard operating procedure (SOP): plan, code, review, test, and human approval. Every task goes through the same steps, which makes code delivery predictable.

AI-processed from GitHub Trending AI; edited by Hamidun News
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceGitHub Trending AI ↗
Editor-in-Chief
Ten CLI agentsCodex, Claude Code, Cursor CLI, OpenCode, and others
SOP stepsagent, shell script, manual approval, nested process
Setupgit clone and start.sh; the first account becomes admin

Anyone who has seriously handed tasks to coding agents knows the problem: the result depends on the particular run. One time the agent writes a plan and tests, another time it skips review, checks its own code, or stops halfway, and the team only finds out in the pull request. For a demo that is tolerable; for team work it is not.

Tencent has published LoopForge, a platform that you install on your own server (self-hosted) and that makes ten coding agents work under a single set of rules. In the project, such a set of rules is called an SOP, a standard operating procedure: a fixed chain of steps such as "plan → code → review → tests → human approval," the same for every new task. The project is distributed under the MIT license, meaning it can be freely used and modified, and it gained about 174 stars on GitHub in a single day.

What it is

LoopForge is a workspace where a team ships software with the help of agents. The agents themselves are not replaced; the platform manages them. The work is carried out by CLIs that are already familiar: CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI, Antigravity, Kimi Code, Pi, DeepSeek Harness, and ZCode. These are the ten agents from the headline. LoopForge adds roles, sandboxes, checkpoints, and run history around them, and everything runs on your infrastructure.

There are two modes. The first is a chat with Main Bot: you describe a task, the platform spins up a Docker sandbox, clones the repositories, delegates work to the project's roles, and opens a PR. This suits exploration, quick fixes, and iteration. The second is the SOP, repeatable delivery with fixed stages and human sign-off. If a way of working in chat has proven itself, it can be saved as an SOP, and the next requirement will follow the same path.

How the process works

An SOP is assembled on a visual canvas from nodes: an agent (a role with its own prompt, executor, model, rules, Skills, and MCP tools), a script, a manual gate, and a nested SOP. A script is an ordinary deterministic shell step: build, tests, linter, report. The authors state explicitly that AI is not used where it is not needed. A manual gate halts the run until a person checks the results and either approves them or sends the work back to a chosen step.

Take a typical bug fix. One role plans the fix, a second writes the code, a third reviews it independently, and a fourth summarizes. If a check fails, the work goes back to the previous step. In the documentation's example, four CodeBuddy roles fixed a Python bug, three real tests passed, the run waited for approval, and the platform verified that the accepted code had not changed afterward. Ready-made SOPs are included for backend, frontend, bug fixing, and product documents.

A run can be paused, steered with messages, or taken over directly from the chat. The platform shows node statuses, tool traces, logs, and result snapshots. A failed step can be retried or restarted from any node. The platform also tracks token spending per SOP and per dialogue at public prices and keeps delivery reports.

Installation and limitations

Deployment uses git clone and the start.sh script. The installer checks for Docker, Compose, Python 3, and curl, brings up MySQL and the platform services, generates secrets, and installs the built-in sandboxes and starter SOPs. The first registered user becomes the administrator. The demo SOP consists only of a script and a manual approval; the chat and agent steps require keys for the chosen executor, since no models are included.

There are several limitations, and they matter. Docker sandboxes run through the host's Docker socket, which amounts to control over the daemon, so the configuration is designed for a trusted team rather than multi-tenancy. New installations listen on all interfaces, colleagues who can see the login screen can register, and signed-in users can see project data across the entire instance. Before exposing it to the internet, you need to put an HTTPS proxy in front of it and disable sign-up with the variable AGENTFLOW_ALLOW_SIGNUP=false. The chat interface is currently only in Chinese; the management pages are in English and Chinese. Some internal features, such as requirement clarification and notification summaries, work only with CodeBuddy.

Why it matters

If your team already uses several coding agents but its workflow depends on individual discipline, it is worth trying LoopForge on the demo SOP and one small bug fix. The practical takeaway is simple: the reliability of agentic development comes not only from a smarter model but also from a rigid framework around it: identical stages, ordinary scripts where AI is not needed, and checkpoints where a human makes the decision.

Story

Key Questions Answered

What it is?

LoopForge is a workspace where a team ships software with the help of agents. The agents themselves are not replaced; the platform manages them. The work is carried out by CLIs that are already familiar: CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI, Antigravity, Kimi Code, Pi, DeepSeek Harness, and ZCode. These are the ten agents from the headline. LoopForge adds roles, sandboxes, checkpoints, and run history around them, and everything runs on your infrastructure. There are two modes.

How the process works?

An SOP is assembled on a visual canvas from nodes: an agent (a role with its own prompt, executor, model, rules, Skills, and MCP tools), a script, a manual gate, and a nested SOP. A script is an ordinary deterministic shell step: build, tests, linter, report. The authors state explicitly that AI is not used where it is not needed. A manual gate halts the run until a person checks the results and either approves them or sends the work back to a chosen step. Take a typical bug fix.

What are the key details on Installation and limitations?

Deployment uses git clone and the start.sh script. The installer checks for Docker, Compose, Python 3, and curl, brings up MySQL and the platform services, generates secrets, and installs the built-in sandboxes and starter SOPs. The first registered user becomes the administrator. The demo SOP consists only of a script and a manual approval; the chat and agent steps require keys for the chosen executor, since no models are included. There are several limitations, and they matter.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…