Tencent/LoopForge

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.

LoopForge — What It Is

LoopForge is a self-hosted workspace for teams that ship software with AI coding agents. You describe what you want in a chat window, and the system coordinates AI agents, sandboxes, and approval steps to turn that request into actual code changes — usually a pull request.

It's built around three ideas:

  1. Chat with a "Main Bot" — you type a request, and it opens an isolated Docker workspace, checks out your repositories, delegates work to specialized AI roles (planner, implementer, reviewer, tester), and opens a PR when done.
  2. SOPs (Standard Operating Procedures) — once you find a workflow that works, you can save it as a repeatable, fixed sequence of steps (AI agents, scripts, and human approval gates). Every future request then follows the same stages with the same checks.
  3. Human control — approval gates pause the run until a person reviews artifacts and approves, sends work back to an earlier step, or retries. Nothing ships without sign-off if you don't want it to.

The actual coding work is done by existing AI coding tools — CodeBuddy, Codex, or Claude Code. LoopForge doesn't replace them; it orchestrates them, adding roles, sandboxes, checkpoints, and history around them. It runs on your own infrastructure and is MIT-licensed.

Two Ways to Work

Chat with Main Bot Run an SOP
Best for Exploring, quick fixes, iterating with the agent Repeatable delivery with fixed stages and sign-off
How it runs Main Bot opens a Docker sandbox on demand and calls project roles as subagents The engine runs Agent, Script, Human-gate, and sub-SOP nodes in order, each in its own sandbox
You stay in control by Queuing or steering messages mid-reply; pausing, resuming, or taking over a running SOP Approving at human gates, sending work back to a chosen step, retrying or rerunning from any node
Delivers A PR/MR, or a direct merge into a shared branch Documents, code, or a PR, plus per-node artifacts and logs

Both paths track the same requirements, so every change links back to what was asked for and counts toward delivery and cost reports.

Why It Exists

Asking a coding agent to "fix this bug" gives results that depend on that particular run — it might skip the plan, forget tests, review its own code, or stop halfway unnoticed. LoopForge addresses this with:

  • Repeatability — same SOP, same stages, same checks for every requirement.
  • Separated roles — planning, implementation, review, and testing each get their own prompt, executor, model, rules, and tools.
  • Human in the loop — approval gates where they matter; you can read the report, approve, or send work back.
  • Observable and resumable — node status, tool traces, logs, and artifacts in one place; pause, retry, or rerun from any node.
  • Your stack, your servers — CodeBuddy, Codex, or Claude Code as executors; GitHub, Skills, and MCP as tools; Docker sandboxes or your own machine as the runtime.

Quick Start

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

The installer checks for Docker, Compose, Python 3, and curl (offering to install missing pieces), pulls a pre-built release image, creates MySQL and platform services, generates secrets, and installs built-in sandboxes, starter SOPs, and a welcome project. You don't need Go or Node.js installed on the host. After startup, you configure model API keys and GitHub authorization.

The service listens on 0.0.0.0:8090 by default. The first account created becomes the administrator. New accounts join a demo project with a script-only SOP (no API key needed) so you can try it immediately.

How SOPs Work

An SOP is the workflow a delivery follows. It's made of these node types:

Node What it does
Agent A role (prompt, executor, model, rules, Skills, MCP tools) that designs, implements, reviews, or tests
Script A deterministic shell step — build, test, lint, generate a report. No AI where AI isn't needed
Human gate Pauses the run until someone approves or sends work back to a step
Sub-SOP Runs another SOP as a sub-workflow
Preview build Builds a static site and archives it for review

Each node declares its artifacts and dependencies; the engine starts each step once its dependencies complete. Built-in SOPs cover backend delivery, frontend delivery, product proposals, and bug fixes — they're read-only blueprints you copy into your project to run and edit.

Key Features

  • Chat workbench — per-conversation executor, model, and permission mode; @ references for roles, skills, requirements, SOPs, and repositories; message queueing and steering; generated files open in a side panel.
  • Conversation workspaces — on-demand Docker sandboxes with repositories checked out and project roles as subagents; pause, resume, or take over a running SOP.
  • SOP designer — visual canvas with all node types, rework-on-failure edges, validation, version history, and publishing.
  • Roles — reusable definitions shared across SOP nodes and chat subagents.
  • Requirement → delivery — end-to-end tracking with PR/MR status, merge detection, and direct-branch merges.
  • Runs & approvals — live node status, tool traces, worker logs, artifact snapshots, and a single "Needs review" inbox.
  • AI scheduled tasks — recurring agent tasks with a shared market of versioned templates.
  • Insights & cost — engineering metrics, business-domain reports, and token/cost accounting.
  • Execution environments — platform workers, Docker sandboxes, or a connected local machine.
  • Team-ready — password accounts, projects and members, light/dark themes, and an OpenAPI-described HTTP and MCP API.

LoopForge Local

Don't want to run a server? LoopForge Local brings the built-in delivery workflows into the coding assistant you already use:

npx -y loopforge-cli@latest install claude

Then run /start-devflow "<your request>" in Claude Code. Supported hosts include Claude Code, CodeBuddy, Codex, Cursor, Pi, DeepSeek Harness, and Hermes Agent. It requires Node.js 20+ and Python 3.8+.

FAQ Highlights

  • No API key needed to try it — the demo SOP uses only a script and a human gate. Chat and agent steps need a key for the executor you pick.
  • GitHub is optional — chat, documents, and new projects run in the task workspace. Connect GitHub when you want to change existing code or open PRs.
  • Different from Dify/n8n/Flowise — those are general-purpose AI workflow builders; LoopForge is built specifically for software delivery with coding agents: implementer/reviewer roles, Docker sandboxes with real repos, GitHub delivery, human acceptance gates, and run evidence.
  • Different from using Claude Code/Codex/CodeBuddy directly — LoopForge orchestrates them; each agent step runs one of those CLIs, and LoopForge adds who does what, in which sandbox, in which order, with which checks, and who signs off.

License

MIT. Third-party components keep their own licenses.

Related

  • Project
  • Project
  • Project
  • Project