LATEST
AI Tools

Tencent LoopForge: Free Coding-Agent Orchestration, With a Human Sign-Off

Updated November 22, 2026
12 min read
Editor's Note: This article was human-reviewed for accuracy
Two developers examining code on a large screen in a modern office

Quick Answer & Key Takeaways

  • • LoopForge is Tencent's free, self-hosted platform for running coding agents through repeatable delivery workflows with human approval gates.
  • • It orchestrates ten coding-agent CLIs, including Claude Code, Codex and OpenCode, but includes no model of its own, so you pay for the agents you plug in.
  • • Version 0.4.0 is stable and MIT-licensed, though the chat workbench is Chinese-only for now and the default setup listens on all network interfaces.

Tencent open-sourced LoopForge on November 16, 2026, and shipped the stable v0.4.0 release the next day. It is a self-hosted workspace where a team ships software with coding agents, and it had collected 5,540 GitHub stars by November 22, up from roughly 2,100 on launch day, according to launch coverage and a live check of the repository. The pitch is simple: a coding agent on its own is unpredictable, so LoopForge wraps the agents a team already uses in repeatable workflows, Docker sandboxes and human approval gates.

What LoopForge is

The important detail is what LoopForge does not do. It writes no code itself and includes no model of its own. It orchestrates ten coding-agent command-line tools, CodeBuddy, Codex, Claude Code, OpenCode, Cursor CLI, Antigravity, Kimi Code, Pi, DeepSeek Harness and ZCode, and adds the roles, sandboxes, checkpoints and history around them. You bring your own API keys or subscriptions, and the platform itself costs nothing under its MIT licence.

LoopForge is the public version of AgentFlow, Tencent's internal platform for the same job, according to launch coverage. The repository was created in July 2026, a smaller local edition came out in September, and the self-hosted platform shipped in early November. That short public history matters: this is a young project with a fast release cadence, and the newest release, v0.4.1-rc.1, arrived on November 20, three days after the stable v0.4.0.

Two ways to work: chat or SOP

LoopForge offers two paths through the same system. In chat, you describe what you need to Main Bot, which opens a Docker sandbox, checks out your repositories, delegates to project roles as subagents, and opens a pull request. In SOP mode, the engine runs a fixed workflow of agent, script, human-gate, sub-SOP and preview-build nodes in order, each in its own sandbox.

The SOP is the idea that sets LoopForge apart from a plain agent harness. A standard operating procedure is a saved sequence of steps: clarify the requirement, design a fix, implement it, review it independently, run the tests as a script, then wait at a human gate for approval. If the review fails, the run sends the work back to an earlier step instead of pushing ahead. Four SOPs ship built in: backend delivery, frontend delivery, product proposal and bug fix. Once a way of working proves itself, a team saves it as an SOP and every later requirement follows the same stages, the same checks and the same sign-off.

Both paths track the same requirements, so every change links back to what was asked for and counts toward delivery and cost reports. Runs show live node status, tool traces, logs and per-node artifacts, and a single Needs review inbox collects every gate waiting on a human. You can pause, steer or take over a running SOP from chat, retry a failed step, or rerun from any node.

The ten executors, and what each one costs

LoopForge does not replace your coding agent. It runs one of ten CLIs per step, and each role in an SOP can pick its own executor, model, rules, skills and MCP tools. The built-in templates still default to CodeBuddy, Tencent's own coding assistant, but the platform runs without a CodeBuddy installation or key once workflows use other CLIs.

ExecutorWhat you payHow it is installed
Claude CodeClaude Pro $20 a month, or $17 a month billed annuallyInstalled from npm on first use, then cached
CodexChatGPT Plus $20 a month; Pro from $100 a monthIn the Docker image
Cursor CLIX Subscription Plus $8 a month, or Premium $30 a month for third-party models at API list priceProvider installer on first chat use
OpenCodeFree and open source; you pay the model providerInstalled from npm on first use, then cached
DeepSeek HarnessDeepSeek API pay-as-you-go, from $0.15 per million input tokens off-peakIn the Docker image
CodeBuddyPersonal key; no public price list foundBundled by default
Kimi Code, Pi, Antigravity, ZCodeYour own key or account with that providerInstalled from npm or on the execution machine

The pricing fine print is the whole point of the table. LoopForge itself is free, but a team of five running Claude Code on Pro pays $100 a month in subscriptions before a single token is billed, and $85 a month if all five pay annually. At November 22, 2026 market rates, $20 is about £15.28 or C$28.69, and $100 is about £76.40 or C$143.43, marked as approximations. The platform also records token and cost accounting for SOPs and conversations at public list prices, so a team can see what each delivery actually burned.

Two rows deserve a closer look. Cursor's pricing page no longer sells standalone Cursor plans; it serves the X Subscription, where Plus at $8 a month covers Grok models only and Premium at $30 a month adds third-party models billed separately at API list price. And DeepSeek's $0.15 rate is the off-peak input price for its V4.1 Flash model; peak hours cost double, and the rate belongs to DeepSeek's API, not to LoopForge.

There is also a server bill. LoopForge needs Docker, Compose, Python 3 and curl, plus a machine to run them on, and a connected local execution machine is supported for teams that would rather run agents on their own hardware. No hosted LoopForge service exists, so there is no per-seat platform fee and no vendor to pay for uptime.

Setting it up

Installation is three commands, and the installer checks for Docker, Compose, Python 3 and curl, offering to install what is missing. It pulls the pinned platform image from GitHub Container Registry, creates the MySQL and platform services, generates configuration and session secrets, and installs the built-in sandboxes, starter SOPs and a welcome project.

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

After startup you create an account, add a model key, connect a repository and choose an SOP. GitHub is optional: chat, documents and new projects run in the task workspace without a repository, and bug fixes still need source code and reproduction details. Connecting GitHub uses the built-in LoopForge OAuth app with a device code, and no personal access token is required. For GitLab, Gitea and other hosts, repositories are checked out over SSH with a personal key.

You can try it without an API key at all. New accounts join a demo project with an SOP made only of a script and a human gate: run it, a real script writes a report, the run waits for approval, and you approve it from the Needs review inbox. Real work needs a key for the executor you pick.

The catches

LoopForge is honest about its rough edges, and the README lists them plainly. The chat workbench is Chinese for now, while the management pages are in English and Chinese. That is the first thing an English-only team will notice, and it limits who can drive the Main Bot chat until the workbench is translated.

The default deployment is built for a trusted team, not for the open internet. New installations listen on all network interfaces on port 8090, the first account created becomes the administrator, and teammates can sign up from the login page unless an administrator sets AGENTFLOW_ALLOW_SIGNUP to false. The README tells you to restrict access to your own IP until the administrator account exists and to put an HTTPS reverse proxy in front before exposing it. SOP runs and conversation workspaces use Docker sandbox containers started through the host Docker socket, and the README states plainly that Docker socket access grants control of the host daemon, so the setup is for trusted teams, not untrusted multi-tenant isolation.

There are no independent benchmarks or hands-on reviews of LoopForge yet. The one end-to-end delivery the README describes, four CodeBuddy roles planning, implementing, reviewing and summarising a Python fix with three real tests passing, is the project's own validation record, and launch coverage describes a 0 to 100 run score without publishing the scoring method. Treat the star count the same way: 5,540 stars in a week is attention, not evidence that the SOPs hold up on a real codebase.

The licence needs one careful note. GitHub's API labels the licence NOASSERTION, because the LICENSE file is Tencent's own MIT wording rather than the standard MIT text. The file itself states that LoopForge is licensed under MIT, except for third-party components that keep their own licences, and the README links the same file. The platform is free to use and self-host, but teams that care about licence compliance should read the file and the third-party notices rather than trusting the badge.

Finally, the release cadence is fast and the project is young. Version 0.4.0 is the current stable release, and v0.4.1-rc.1 followed three days later as a prerelease. Early adopters should expect frequent upgrades and should read the changelog and the migration files before pulling a new version.

How LoopForge compares

LoopForge sits between single-agent runtimes and hosted coding services. It is not a model, not a hosted agent and not a general automation builder. The comparison below uses each tool's own positioning, and the prices are the ones that apply to a team that self-hosts.

ToolWhat it isPriceAPI access
LoopForgeSelf-hosted SOP orchestration for ten coding-agent CLIsFree, MIT licence; you pay the agentsOpenAPI HTTP and MCP API
OpenClawLocal-first personal assistant runtime you control from chat appsFree, MIT licence; you pay the modelYes, local runtime
MyClaw 3.0Hosted AI teammates on any harness, with creditsPro $20 a monthVia its own platform
Claude CodeOne coding agent in the terminal, no orchestration layerClaude Pro $20 a monthYes, Agent SDK
Eclipse Enclave 1.0Free sandbox for AI coding agentsFree, self-hostedYes
DevinHosted autonomous coding service with desktop and botsPro $20 a month, Max $200 a monthYes

The honest way to read that table is as a trade, not a ranking. LoopForge gives a team repeatable process and a sign-off step, and charges nothing for it, but it also hands the team the server, the security model and the translation gap. Devin and MyClaw charge a subscription and take the hosting away. OpenClaw costs nothing and runs locally, but it is a personal assistant runtime, not a team delivery workflow. Our roundup of the 7 best AI coding agents in 2026 covers the per-agent prices and limits in more detail.

Who should try it, and who should wait

LoopForge fits a small team that already pays for one or two coding agents and wants every delivery to follow the same steps, with a human signing off before anything merges. The demo SOP runs without a key, so the cost of finding out is an afternoon on a spare machine, not a subscription. Teams that ship in Chinese, or that already run Docker infrastructure, will feel at home fastest.

Wait if your team needs an English chat workbench today, if you want a vendor-hosted service with support, or if you need published evidence that the SOPs improve delivery quality. None of those exist yet. Also wait if you cannot give the platform a trusted network: the default bind and open sign-up are the two settings to change before the service leaves your own machine.

Common mistakes

First, exposing the default install. The service listens on 0.0.0.0:8090 and the first account becomes admin, so create that account and lock down access before the port is reachable beyond your team. Second, assuming the platform is free in full. The licence is free; the agents are not, and five Pro seats plus a server is a real monthly bill. Third, treating the star count as a review. LoopForge gained about 700 stars on launch day and passed 5,500 within a week, which shows interest, not maturity. Fourth, running untrusted code in the sandbox and assuming isolation. The README is explicit that Docker socket access is for trusted teams. Fifth, skipping the changelog on upgrade. Migrations ship as files under migrations/, and v0.4.1-rc.1 already includes a credential migration that needs a manual step after verification.

Frequently asked questions

Is LoopForge really free?

Yes, the platform itself is free under the MIT licence, and there is no paid tier or per-seat platform fee. You pay for the coding agents you plug in, such as Claude Code on Claude Pro at $20 a month, plus the server you run it on.

Does LoopForge include its own AI model?

No. LoopForge writes no code itself. It orchestrates ten coding-agent CLIs, including Claude Code, Codex, OpenCode and CodeBuddy, and each step uses the executor and model you configure.

What do I need to run it?

Docker, Compose, Python 3 and curl, plus a machine to run them on. The installer checks for each and offers to install what is missing. macOS needs Docker Desktop and a working Python 3.

Can I try it without an API key?

Yes. New accounts join a demo project with an SOP made only of a script and a human gate. It writes a real report and waits for your approval, with no key required. Chat and agent steps need a key for the executor you pick.

Is it safe to put on the internet?

Not with the defaults. New installations listen on all interfaces, the first account becomes admin, and sign-up is open unless you turn it off. Restrict access to your own IP, create the admin account first, and put an HTTPS reverse proxy in front before exposing it.

What is the difference between LoopForge and LoopForge Local?

The platform is the team workspace with chat, SOPs, approvals and shared history, and it needs Docker. LoopForge Local is the legacy single-developer edition that runs built-in workflows inside a coding assistant with Node.js 20 or newer and Python 3.8 or newer.

Verdict

LoopForge is a promising answer to a real complaint: a coding agent on its own does not follow a process, and a team cannot ship on one-off chats. Tencent gives the orchestration away for free, supports ten agents instead of locking you to one, and puts a human gate between the agent and the merge. The price is that you host it, secure it, and read the chat in Chinese for now. For a team that already pays for coding agents and wants repeatable delivery, that trade is worth an afternoon of testing. For everyone else, the demo SOP costs nothing and tells you quickly whether the workflow fits.

Sources

Exchange rates: 1 USD = 0.764 GBP and C$1.434, November 22, 2026 market rates, verified at publication and marked as approximations.

12 min read
11 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more