Tencent released LoopForge, a self-hosted MIT-licensed platform that orchestrates ten coding-agent CLIs — CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI, Antigravity, Kimi Code, Pi, DeepSeek Harness, and ZCode — into repeatable software-delivery workflows. The system centers on two paths: an interactive chat with Main Bot that spins up on-demand Docker sandboxes, checks out repositories, delegates to specialized roles (planning, implementation, review, testing), and opens PRs; and SOPs (Standard Operating Procedures) that codify proven workflows as directed acyclic graphs of Agent, Script, Human-gate, Sub-SOP, and Preview-build nodes. Each role binds its own prompt, executor, model, rules, Skills, and MCP tools, and runs in an isolated sandbox with full tool traces, logs, and artifact snapshots. Human gates enforce approval, rework, or retry from any node. Installation via ./start.sh pulls a pinned GHCR image (configurable via MIRROR ), provisions MySQL and platform services, seeds built-in SOPs (backend, frontend, bug-fix, product), and listens on 0.0.0.0:8090; it requires Docker, Compose, Python 3, and curl, supports Linux (apt/dnf/yum), macOS (Docker Desktop), and validates Compose plugin SHA-256. GitHub OAuth (built-in app) or PAT enables PR/MR automation; GitLab, Gitea, and Gogs work via SSH keys. The platform tracks requirements end-to-end across chat and SOP runs, exposes an OpenAPI/MCP HTTP API, and provides cost accounting at public list prices. LoopForge Local packages the same built-in workflows as an npx -installable CLI for Claude Code, CodeBuddy, Codex, Cursor, Pi, DeepSeek Harness, and Hermes Agent (Classic and Portable editions), storing artifacts under artifacts/{task_slug}/ without a server.

Read original