Image: GitHub
Why it mattersA team running coding agents can now pin every change to a checked procedure. The run stops at each gate until a named person signs it off.
Tencent released LoopForge, a self-hosted platform for coding-agent work. The repository has 973 stars. The README calls it "self-hosted workflows for coding-agent delivery". It is built for teams that deliver software with coding agents. A single developer running one CLI is already too small a unit.
What it does
A user talks to the platform in chat, called Main Bot. Main Bot opens a Docker sandbox and checks out the team's repository. It hands the work to the right role. It opens a pull request at the end. The same platform also runs Standard Operating Procedures. An SOP is a saved workflow with fixed stages, scripts, agents and human gates. Tencent included four templates: a backend build, a frontend build, a bug fix, and a product proposal. The bug-fix template has planning, implementation, an independent review, and testing. The product-proposal template writes a document. Every step is linked to the request that started it.
The agents it drives
LoopForge is an orchestrator. It picks a coding-agent CLI for each step. The list has ten names. They are CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI, Antigravity, Kimi Code, Pi, DeepSeek Harness and ZCode. A user chooses the executor and the model for each step. Tencent says the point is to "orchestrate" these CLIs and leave the CLIs themselves in place. The SOP designer is a visual canvas. Its nodes are Agent, Script, Human gate, Sub-SOP and Preview. A failed step can go back to an earlier node for rework.
How the gates work
A human gate pauses the run. The reviewer has three choices. They can approve and continue. They can add extra context and run the same step again. They can send the work back to an earlier step. A single inbox lists every gate that is waiting on a person. Tencent also added a demo project. It walks a reader through a script-and-approval workflow with no API key. A team can try the gate pattern before handing the platform any credentials.
How to run it
LoopForge is MIT-licensed. It is a Docker Compose service with a web interface. It runs MySQL for state. It opens a Docker sandbox for each run. Tencent also published a second build, LoopForge Local. That version runs inside the coding assistant itself, with no separate server. It works with Claude Code, CodeBuddy, Codex, Cursor, Pi, DeepSeek Harness and Hermes Agent.
Teams that already run coding agents often let each developer pick their own CLI. The run is not written down. Two developers get different results from the same request. Nothing ties the change back to who approved it. LoopForge gives each requirement the same workflow. It forces the review to happen at a named step. It keeps the evidence next to the pull request. The ten-CLI list means a team can keep whichever assistant each developer prefers. The executor can change per step, without rebuilding the workflow.
Source
Primary source: Tencent/LoopForge on GitHub.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.