Rank #07 · Daily · DevTools
LoopForge: Tencent's Self-Hosted Workflows for Coding-Agent Delivery
Describe a change to Main Bot, or run a saved SOP of agents, scripts, and human gates, and ship the work from a Docker sandbox on your own infrastructure.
What it does
LoopForge is a self-hosted workspace from Tencent for shipping software with coding agents. GitHub description: "Self-hosted workflows for coding-agent delivery. Orchestrate CodeBuddy, Codex and Claude Code with scripts and human approval gates. Includes Local workflows for coding assistants." Site: loopforge.cn.
README and LICENSE (as verified):
- Two working paths share the same requirements (and delivery / cost reports). Chat with Main Bot opens a Docker sandbox, checks out repositories, delegates to project roles, and can open a PR (or merge directly with
mode=direct). Run an SOP — a standard workflow of Agent, Script, Human-gate, and sub-SOP nodes — executes those stages in their own sandbox, with approval, retry, and send-back controls. - Coding-agent CLIs do the implementation work. The README lists CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI, Antigravity, Kimi Code, Pi, DeepSeek Harness, and ZCode. LoopForge adds roles, sandboxes, checkpoints, and history around them.
- SOP designer: a visual canvas with Agent, Script, Human gate, Sub-SOP, and Preview-build nodes. Built-in (read-only) blueprints cover backend, frontend, bug-fix, and product delivery; copy one into a project to edit it. Roles reuse a prompt, executor, model, rules, Skills, and MCP tools across SOP nodes and chat subagents.
- Human gates pause a run until someone reviews artifacts and approves, adds context, or sends the work back to a chosen step. Runs expose node status, tool traces, logs, and artifacts; a Needs review inbox collects pending gates.
- Quick start (after installing Git): clone the repo and run
./start.sh, which checks Docker, Compose, Python 3, and curl, then pulls the GHCR image pinned incompose.yaml. It starts MySQL and platform services, generates secrets, and installs built-in sandboxes, starter SOPs, and a welcome project. Configure model API keys and optional GitHub authorization after startup. Host Go and Node.js are not required for this path. The demo SOP (script + human gate) does not need an API key; chat and agent steps do. Model usage is not bundled — you bring credentials. - Default Compose installs start SOP and conversation workspaces as Docker sandbox containers via the host Docker socket. The README warns that socket access grants control of the host daemon and is intended for trusted teams, not untrusted multi-tenant isolation. New installs listen on all interfaces; signed-in users can read project data across the instance unless you tighten the account model (
AGENTFLOW_ALLOW_SIGNUP=falseturns off open signup). - LoopForge Local (
npx -y loopforge-cli@latest install claude, then/start-devflowin Claude Code) runs built-in workflows inside an existing coding assistant, without the platform server. Classic and Portable editions cover Claude Code, CodeBuddy, Codex, and Cursor; Portable also lists Pi, DeepSeek Harness, and Hermes Agent. Exporting a platform SOP for standalone local execution is described as planned, not available today. - Other documented pieces include AI scheduled tasks, insights and token/cost accounting, an OpenAPI HTTP and MCP API, and English/Chinese management pages. The README notes that the chat workbench is Chinese for now.
License: this repository's own code is MIT (LICENSE copyright 2026 Tencent; README badge and license section agree). Third-party components retain their original licenses and notices — see LICENSE and THIRD_PARTY_NOTICES.md. GitHub's license API currently shows NOASSERTION because of that mixed file; do not treat the badge as "unlicensed." This is a summary, not legal advice.
Who it's for
- Teams that want self-hosted delivery workflows around coding agents they already use, with review and approval built in.
- Groups that want a repeatable SOP (fixed roles, tests, and sign-off) once a chat-driven path has proven itself.
- Individual developers who prefer LoopForge Local inside Claude Code, Codex, Cursor, or another listed host, without running the full platform.
Less ideal if you only need a single coding-agent CLI with no shared workspace or gates, cannot run Docker (for the platform path), need the chat workbench in English today, or want a multi-tenant SaaS with hard isolation — the default socket-based setup is for a trusted team.
Stars / language / license
3,148 stars · Go · MIT
Data verified: 2026-11-18 09:02 Asia/Shanghai (official Daily Top10 gate; language confirmed via GitHub API get_repository; MIT plus third-party license exception confirmed via LICENSE + README; API SPDX shows NOASSERTION)
Trending note: +591 stars today (captured 09:02 Asia/Shanghai)
About the license: MIT covers LoopForge's own code. Third-party components keep their original terms. This is a summary, not legal advice.
Category tags
DevTools · workflow · coding-agents · self-hosted · Go · MIT · human-in-the-loop · Tencent
How it differs from similar projects
Compared with running a coding-agent CLI alone. LoopForge's pitch is the orchestration layer: sandboxed workspaces, named roles, saved SOPs, human gates, and a record that ties work back to a requirement. The agent CLI still writes the code.
Compared with Plannotator (#05). Plannotator is a local review and annotation surface that sends your comments back to an agent session. LoopForge is a delivery workspace that runs agents, scripts, and approval gates to produce a PR or artifact. They can sit in the same loop; TidyHow has not compared them as substitutes.
Compared with Open Code Review (#04). That project generates line-level review comments. LoopForge can include review as an SOP role or gate, but it is a workflow platform, not a dedicated review CLI.
Compared with a general CI system (GitHub Actions and similar). CI is typically pipeline-as-code on a forge. LoopForge is a chat-plus-SOP workspace with agent executors and in-product human gates. It can open PRs into your existing forge; it is not presented as a replacement for CI.
Caveats. Default Docker-socket deployment is for trusted teams. Chat and agent steps need your own model credentials. The chat workbench is Chinese for now. Author-documented sample deliveries and validation notes are not success-rate benchmarks; no independent workflow success figures are claimed here.