What is LoopForge?
LoopForge is a self-hosted web workspace for shipping software with coding-agent CLIs. You can chat with a Main Bot that opens a Docker sandbox, checks out your repositories, hands work to roles such as planner, implementer and reviewer, and opens a pull request; or you can run an SOP, a versioned workflow of agent, script, human-approval and sub-workflow steps drawn on a visual canvas.
It drives ten agent CLIs, including CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI and Kimi Code, and tracks requirements, run logs, artefacts and token costs. Installation is a Docker Compose stack with MySQL, and a separate LoopForge Local package puts the built-in workflows into Claude Code, Codex, Cursor and other assistants without a server. GitHub shows the licence as NOASSERTION because the LICENSE file wraps the MIT text in a Tencent preamble; the terms are MIT, and the one bundled third-party component, a modified copy of obra/superpowers, is also MIT.
Is LoopForge worth it?
LoopForge's argument is that a single agent run is not a delivery process, and its design follows through. Each SOP step gets its own role, executor, model, rules and tools, so the reviewer is not the agent that wrote the code; script steps run builds and tests without AI; and human gates pause a run until someone approves or sends the work back to a chosen step. Built-in SOPs cover backend, frontend, bug-fix and product work. The README records one delivery in which four CodeBuddy roles planned, implemented and independently reviewed a Python fix, three tests passed and the run waited for human approval.
It is candid about its origins. Tencent published it in July 2026, and version 0.4.0 shipped on 17 November with fixes that keep credentials out of clone URLs and confine agent configuration writes. Its public-scope document describes an adaptation of an internal platform, with sync records kept private and internal tracker code hidden by default. CodeBuddy, Tencent's own agent, is the default for built-in roles, and requirement clarification, notification summaries, Bot replies and run evaluations use CodeBuddy models only. The chat workbench is in Chinese for now; management pages are in English and Chinese.
The catch: the security model assumes everyone is trusted. A new install listens on 0.0.0.0:8090, anyone who reaches the login page can sign up until `AGENTFLOW_ALLOW_SIGNUP=false` is set, and every signed-in account can list all projects, requirements and runs. The Compose setup mounts the host Docker socket, which the docs call equivalent to root on the host, so anyone who can run an SOP can execute code on that machine. Agent CLIs also obey a bound repository's own `CLAUDE.md`, `.claude/settings.json` hooks and `.mcp.json`, so an untrusted repository can start commands inside the sandbox. Model keys are stored per user and model usage is not included.
Our take: for an engineering team that already uses two or three coding agents and wants the same review and approval path on every change, LoopForge is a more complete answer than scripting one yourself, provided it runs behind a VPN or reverse proxy with sign-up off. Archon covers similar ground if you would rather write workflows in YAML than draw them on a canvas.
Where it earns its place
A ten-person product engineering team using Claude Code and Codex
Make every agent-written change go through planning, independent review, tests and a human sign-off
Copy the built-in backend or bug-fix SOP into a project, set each role's executor and model, and connect GitHub through the built-in OAuth device flow. Runs wait in a Needs review inbox before a PR goes out. The software is free; model usage is billed on your own keys.
A solo developer in Claude Code or Cursor
Get a structured plan, implement and review flow without running a server
Run `npx -y loopforge-cli@latest install claude` inside the project, then `/start-devflow "<request>"`. Needs Node.js 20+ and Python 3.8+; artefacts land under `artifacts/{task_slug}/`.
An engineering manager reporting on agent use
See what each requirement cost in tokens and how it was delivered
Insights pages report engineering metrics and token and cost accounting per SOP and conversation at public list prices, with static HTML and SVG export.
Before you commit
- Setup
- Clone the repository and run `./start.sh`; it pulls the pinned GHCR image, creates MySQL and platform containers, generates secrets and seeds starter SOPs. The first account becomes administrator. A script-and-approval demo SOP runs without any API key.
- Integrations
- CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI, Antigravity, Kimi Code, Pi, DeepSeek Harness and ZCode as executors; the GitHub API for PRs and issues; GitLab, Gitea and others over Git SSH for clone and push only; Skills and MCP tools; an OpenAPI-described HTTP and MCP API.
- Data and privacy
- Self-hosted, with data in MySQL and a local data directory. Passwords are bcrypt-hashed and sessions expire after seven days. Code and prompts go to whichever model provider each executor uses. Public defaults install no enterprise MCP servers and contact no corporate identity services, per the public-scope document.
- Cost at scale
- Free under MIT; you pay for the server and for model usage on each user's key or a shared key. Token costs are tracked per run at public list prices.
- Limits
- Built for a trusted team, not isolation between untrusted tenants; GitHub is the only host with built-in API automation; Compose conversation workspaces use Docker sandboxes only; exporting platform SOPs to Local is planned, not shipped.
What’s good
- Separate roles per step, script checks and human approval gates in one workflow
- Ten agent CLIs as executors, chosen per role
- Requirement tracking, run logs, artefacts and token cost reports
- MIT licence; LoopForge Local brings the workflows to a single developer without a server
What’s not
- Listens on all interfaces with open sign-up by default
- The Docker socket mount gives anyone who can run an SOP root-equivalent access to the host
- Several features use CodeBuddy models only; the chat workbench is Chinese for now
- Young: first published in July 2026, now at version 0.4.0
Best for
Engineering teams of trusted colleagues who want coding agents to follow the same plan, review, test and sign-off steps on every change, on their own servers.
LoopForge alternatives
LoopForge: common questions
- Is LoopForge worth it?
- LoopForge gives coding agents a real delivery process, with separate reviewers and human sign-off, but its defaults suit a trusted team on a private network and nothing wider. We rate it 3.7 out of 5.
- How much does LoopForge cost?
- Free and open source under the MIT licence; model usage is billed by each provider on your own keys.
- What are the downsides of LoopForge?
- Listens on all interfaces with open sign-up by default; the Docker socket mount gives anyone who can run an SOP root-equivalent access to the host; several features use CodeBuddy models only; the chat workbench is Chinese for now; young: first published in July 2026, now at version 0.4.0.
- Who is LoopForge best for?
- Engineering teams of trusted colleagues who want coding agents to follow the same plan, review, test and sign-off steps on every change, on their own servers.
- How much does LoopForge cost for a team or in production?
- Free under MIT; you pay for the server and for model usage on each user's key or a shared key. Token costs are tracked per run at public list prices.
- Is LoopForge safe for company data?
- Self-hosted, with data in MySQL and a local data directory. Passwords are bcrypt-hashed and sessions expire after seven days. Code and prompts go to whichever model provider each executor uses. Public defaults install no enterprise MCP servers and contact no corporate identity services, per the public-scope document.
- What are the best LoopForge alternatives?
- The closest alternatives we have reviewed are Archon, Paperclip and OpenHands.
Answers checked Nov 18, 2026.