Last updated: November 15, 2026
Tencent LoopForge is an open-source, self-hosted platform, announced November 9, 2026, that runs coding-agent CLIs such as Claude Code, Codex and CodeBuddy inside Docker sandboxes. Teams chat with a Main Bot or run SOPs with agent, script and human-approval steps, and the latest release is v0.3.9-rc.1.
This guide draws on the LoopForge GitHub README, the release notes and Tencent’s launch article on the Tencent Cloud developer community. If you use a single coding agent today, our Codex CLI 0.169 guide and Claude Code Code Mode guide cover the tools LoopForge orchestrates.
Contents
- What is new
- Chat or SOP: two ways to work
- How to install the platform
- Executors and credentials
- LoopForge Local
- Recent releases
- Security and limits
- Frequently asked questions
What is new
Tencent’s Lighthouse (lightweight cloud) team published LoopForge as open source on November 9, 2026. According to the launch article, the platform was already used by more than 20 internal business teams and completed more than 200 requirements and routine tickets automatically in a single month. Tencent pitches it as a fix for team problems with single-machine agents, such as work stopping when a laptop closes.
The README describes the core idea in one line: describe a requirement to Main Bot, which opens a Docker sandbox, checks out your repositories, delegates to project roles and opens the PR. When a way of working proves itself, you save it as an SOP so the next requirement follows the same stages and checks.
Chat or SOP: two ways to work
| Chat with Main Bot | Run an SOP | |
|---|---|---|
| Best for | Exploring, quick fixes, iterating | Repeatable delivery with fixed stages and sign-off |
| How it runs | Opens a conversation workspace (a Docker sandbox with your repos) and calls project roles as subagents | Runs Agent, Script, Human-gate and Sub-SOP nodes in order, in their own sandbox |
| Control | Queue or steer messages mid-reply; pause, resume or take over a running SOP | Approve at gates, send work back to a chosen step, retry or rerun from any node |
| Delivers | A PR/MR, or a direct merge into a shared branch (mode=direct) |
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. An SOP can contain five node types: Agent, Script, Human gate, Sub-SOP and Preview build. Built-in backend, frontend, bug-fix and product SOPs are included.
Roles keep work separated: planning, implementation, review and testing each get their own prompt, executor, model, rules, Skills and MCP tools, so the agent that writes code does not review it.
How to install the platform
The platform runs as Docker containers on your own server. The installer checks for Docker, Compose, Python 3 and curl and offers to install anything missing.
- Clone the repository:
git clone https://github.com/Tencent/LoopForge.git, thencd LoopForge. - Start it: run
./start.sh. It pulls the release image pinned incompose.yamlfrom GHCR, creates the MySQL and platform services, generates secrets and installs the starter SOPs and a welcome project. SetLOOPFORGE_YES=1to accept prompts automatically. - Create the admin account: open the printed address (port 8090 by default). The first account becomes the administrator, so restrict access to your own IP until it exists.
- Try the demo: in the demo project, open Templates, then Platform check ยท script & approval, click Run SOP and approve it from Needs review. No API key is needed.
- Add credentials and a repository: add a model key for your chosen executor, then use Connect GitHub in personal settings, which uses a device code instead of a personal access token.
Manage the service with ./start.sh logs, stop, restart or pull. A MIRROR variable helps when ghcr.io is slow. GitLab, Gitea and Gogs work over Git SSH, but PR/MR automation for them needs a hosting API token or a separately configured MCP.
Executors and credentials
LoopForge does not replace coding agents; each agent step and chat turn runs one of ten CLIs. The README lists how each one is provided:
| Executor | Credentials | Installation |
|---|---|---|
| CodeBuddy | Personal key, or a shared admin key | In the Docker image |
| Codex | Personal key; a connected machine can use its local login | In the Docker image |
| Claude Code | Personal key; a connected machine can use its local login | From npm on first use |
| OpenCode, Kimi Code, Pi | Personal key only | From npm on first use |
| DeepSeek Harness | Personal key only | In the Docker image |
| Cursor CLI, Antigravity | Personal key only | Provider installer on first chat use; not in SOP sandboxes |
| ZCode | Personal key only | Not bundled; install on PATH |
Built-in roles default to CodeBuddy. For background on some of these executors, see our OpenCode background subagents guide and DeepSeek Harness Creator Mode guide.
LoopForge Local
If you do not want to run a server, LoopForge Local installs the built-in delivery workflows into an assistant you already use. Inside your project, run npx -y loopforge-cli@latest install claude, then start a run with /start-devflow "<your request>" in Claude Code. Claude Code, CodeBuddy, Codex and Cursor support the Classic and Portable editions; Pi, DeepSeek Harness and Hermes Agent support the Portable edition. Local needs Node.js 20+ and Python 3.8+, and stores evidence under artifacts/{task_slug}/. Exporting a custom platform SOP to Local is listed as planned, not available. Our Hermes Agent 0.23 guide covers one of the supported hosts.
Recent releases
The releases page shows a fast cadence since launch: v0.3.4 on November 9, v0.3.5 and v0.3.6 on November 11, v0.3.8 on November 12 and the v0.3.9-rc.1 pre-release on November 13, 2026.
- v0.3.8: renders sandbox dev servers in the side browser, asks before handing a requirement to an SOP, and stops trusting identity headers from forwarded clients.
- v0.3.9-rc.1: installs a missing CLI from a failed conversation, supports SSH repositories with per-host user keys, lets you edit run workspaces while agents run, bundles DeepSeek Harness and installs agent CLIs in SOP sandboxes, and keeps credentials out of clone URLs.
To upgrade, check out the new tag, run ./start.sh and apply new files under migrations/.
Security and limits
- Trusted teams only. Sandboxes start through the host Docker socket, which grants control of the Docker daemon. The README says this setup is not meant for untrusted multi-tenant isolation.
- Open by default. New installs listen on all interfaces, signed-in users can read project data across the instance, and teammates can sign up unless you set
AGENTFLOW_ALLOW_SIGNUP=false. Add HTTPS before exposing it. - Bring your own models. Model usage is not included. Insights report token cost at public list prices.
- Language. Management pages are in English and Chinese, but the chat workbench is Chinese only for now.
Compared with Devin’s hosted approach in our Devin Apps guide, LoopForge keeps everything on your own infrastructure, at the cost of running and securing it yourself.
Frequently asked questions
What is Tencent LoopForge?
LoopForge is a self-hosted workspace from Tencent for shipping software with coding agents. A Main Bot in chat opens Docker sandboxes, checks out repositories and opens pull requests, and repeatable work can be saved as an SOP with agent, script and human-approval steps.
Is LoopForge free and open source?
Yes. Tencent open-sourced LoopForge on GitHub, and the README states the MIT license. Model usage is not included, so you bring your own API keys for the coding-agent CLI you choose.
Which coding agents does LoopForge support?
The README lists ten executor CLIs: CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI, Antigravity, Kimi Code, Pi, DeepSeek Harness and ZCode. Built-in SOP roles default to CodeBuddy, and each role can pick its own executor.
Can I try LoopForge without an API key?
Yes. New accounts join a demo project with an SOP made only of a script and a human gate. You can run it, approve it from Needs review and see the final artifact without any model key.
What is LoopForge Local?
LoopForge Local brings the built-in delivery workflows into a coding assistant you already use, with no server. You install it with npx loopforge-cli, then start a run with /start-devflow in Claude Code. It needs Node.js 20+ and Python 3.8+.
Sources
- Tencent/LoopForge README, GitHub
- LoopForge releases, GitHub (v0.3.4 to v0.3.9-rc.1)
- LoopForge open-source launch article, Tencent Cloud Developer Community (November 9, 2026)
- LoopForge official site