Post

LoopForge: Structured Workflows for Coding Agents

Build repeatable coding-agent workflows with review gates, scripts, and self-hosted execution.

LoopForge: Structured Workflows for Coding Agents

A coding agent can make a change quickly, but a team still needs to know what ran, whether tests passed, and who approved the result. LoopForge puts those steps into a self-hosted workspace. It wraps coding-agent command-line tools in roles, sandboxed workspaces, scripts, and human checkpoints. Its source is available under the MIT license.

The useful idea is not that an agent can write code. It is that a team can make the route from request to reviewed change explicit, then reuse that route for the next task.

Two ways to run work

LoopForge’s main interface offers chat with a coordinator called Main Bot or a repeatable workflow called an SOP. In chat, the bot can open a Docker workspace, use project roles, and prepare a change. An SOP defines the stages in advance: agent nodes for work such as design or review, script nodes for commands, human gates for decisions, and sub-workflows for reusable sequences. Nodes can declare dependencies and artifacts, and a reviewer can send work back to an earlier step. The project README describes both paths and their controls.

For a bug fix, a sensible SOP might collect reproduction details, ask one role to implement the fix, run a deterministic test script, and pause for a person to inspect the diff and results. A separate review role gives the team another perspective; it does not guarantee that a defect will be found. A passing script proves only what its tests actually cover.

The platform records node status, logs, and artifacts, and supports retrying or rerunning steps. That history is useful when a run stops halfway: inspect the failed stage and its inputs before deciding whether to retry. Keep acceptance criteria in the task itself so the human gate has something concrete to check.

Start with the no-key demo

The installation guide describes a Docker Compose installation. It requires Git, Docker and Compose, Python 3, curl, and network access; the installer may offer to install missing dependencies. Review the installer and deployment settings before running it on a machine you control:

1
2
3
git clone https://github.com/Tencent/LoopForge.git
cd LoopForge
./start.sh

The installer pulls the image pinned by the checkout, starts the platform and MySQL services, and prints an address. The first account created becomes administrator. The repository recommends using a release tag so the checkout’s configuration and image match. Its built-in demo runs a script followed by a human approval gate without a model API key. Open the demo project, find Platform check · script & approval under Templates, run it, inspect the report, and approve the waiting step. This exercises the workflow mechanics before connecting a model or repository.

Real agent work needs credentials for the selected executor. Existing-repository changes also need repository access. Model calls and infrastructure can incur costs; self-hosting the orchestrator does not make a model provider free. LoopForge supports several executors, including Codex and Claude Code, but their installation and authentication requirements differ. Check the current documentation for the executor and environment you choose.

Understand the trust boundary

The default service listens on all interfaces, and the first browser account becomes administrator. Restrict network access before starting setup; for local-only use, the deployment guide documents LOOPFORGE_BIND=127.0.0.1. It also documents disabling open signup with AGENTFLOW_ALLOW_SIGNUP=false. Signed-in users can view project data across the instance, so this is a trusted-team service rather than a boundary between unrelated tenants.

The default Compose installation gives the platform access to the host Docker socket. The maintainers warn that this is equivalent to root access on the host. Although individual tasks run in containers, users allowed to run workflows can execute code there, and repository configuration can run commands or MCP servers inside a sandbox. Connect trusted repositories, limit accounts and network access, and read the deployment guide’s account and sandbox sections before exposing the service. A human approval gate is a workflow control, not a substitute for those deployment controls.

When it fits

LoopForge is most useful when a team repeats a delivery process and wants explicit handoffs and an audit trail. For a one-off local change, the extra server, database, credentials, and workflow maintenance may be more machinery than necessary. The project also documents a local workflow option that runs inside a supported coding assistant without starting the platform; compare its capabilities with the server before choosing a deployment.

This post is licensed under CC BY 4.0 by the author.