skulldc

LoopForge runs coding agents through self-hosted SOPs with human approval gates

Tencent/LoopForge wraps ten coding-agent CLIs in Docker sandboxes, role-based workflows and approval gates, which matters for teams that want repeatable agent delivery.

LoopForge, a Go project published under the Tencent organization on GitHub, describes itself as a self-hosted workspace where a team ships software with coding agents. The repository was created on 2026-07-28 and had 1383 stars at the time of the trending snapshot, 484 of them gained in the last day. The README states the project is under the MIT license, while the GitHub metadata lists the license as NOASSERTION.

The pitch is aimed at a specific weakness. According to the README, asking a coding agent to fix a bug gives a result that depends on that run: the agent might skip the plan, forget the tests, review its own code or stop halfway without anyone noticing. LoopForge adds roles, sandboxes, checkpoints and history around the agents instead of replacing them.

Two ways to work

The first path is a chat with Main Bot. A request typed into chat makes Main Bot open a Docker sandbox, check out the repositories, delegate to project roles and open the PR. Users can queue or steer messages while a reply runs, and pause, resume or take over a running SOP from chat.

The second path is an SOP, a standard workflow of agents, scripts and human gates. The engine runs its nodes in order, each in its own sandbox:

  • Agent: a role with its own prompt, executor, model, rules, Skills and MCP tools that designs, implements, reviews or tests.
  • Script: a deterministic shell step such as build, test, lint or report generation.
  • Human gate: pauses the run until someone checks the artifacts and approves, or sends the work back to a chosen step.
  • Sub-SOP and Preview build: run another SOP as a sub-workflow, or build a static site for review.

Built-in backend, frontend, bug-fix and product SOPs are included as starting points. Both paths track the same requirements, so every change links back to what was asked for and counts toward delivery and cost reports, with token and cost accounting at public list prices.

Executors and setup

The actual work is done by existing coding-agent CLIs: CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI, Antigravity, Kimi Code, Pi, DeepSeek Harness and ZCode. Each SOP role can pick its executor independently; built-in roles default to CodeBuddy. The README notes that requirement clarification, notification summaries, Bot replies and run evals currently use CodeBuddy only.

The quick start needs Git, and the installer checks for Docker, Compose, Python 3 and curl:

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

According to the README, the script pulls the GHCR release image pinned in compose.yaml, creates the MySQL and platform services, generates configuration and session secrets, and installs the built-in sandboxes and starter SOPs. Host Go and Node.js installations are not required. A demo SOP made only of a script and a human gate runs without any API key.

Security boundaries to read first

The README is direct about the trust model. The service listens on 0.0.0.0:8090 by default, and the first account created becomes the administrator, so the README advises restricting access to your own IP until that account exists. In the default Compose installation, sandboxes start through the host Docker socket, and Docker socket access grants control of the host daemon. The project says this setup is for trusted teams, not untrusted multi-tenant isolation.

Signed-in users can read project data across the instance, and teammates can sign up from the login page unless AGENTFLOW_ALLOW_SIGNUP=false is set in .env. Anyone planning to expose an instance beyond one team is pointed to the account and access model in the deployment docs.


Source: Tencent/LoopForge on GitHub