DEV Community

Cover image for Tencent LoopForge: Turn Coding Agents Into Repeatable Software Delivery
Sh Raj
Sh Raj

Posted on Edited on Fully Autonomous

Tencent LoopForge: Turn Coding Agents Into Repeatable Software Delivery

A coding agent can write a patch quickly. Getting that patch through planning, tests, review, approval and delivery is a separate engineering problem.

Tencent LoopForge is a self-hosted workspace for orchestrating coding-agent CLIs around repeatable software-delivery workflows. It does not replace Claude Code, Codex or CodeBuddy. It adds the process around them: project roles, Docker sandboxes, workflow steps, run history and human approval gates.

LoopForge appeared at #4 on the November 18, 2026 GitHub daily Trending archive. Its v0.4.0 release shipped on November 17. That is a fresh adoption signal, not proof of production maturity.

LoopForge project overview from the official repository

Official project artwork from the LoopForge repository.

Why another coding-agent platform?

A typical agent session ends with a diff and a message saying the work is complete. A team still needs to know whether tests actually ran, whether a separate reviewer inspected the change, who approved it and where the evidence lives.

You can connect scripts, CI jobs and chat tools yourself, but that process is difficult to reuse consistently. LoopForge makes the workflow explicit.

How it works

There are two main paths.

Main Bot chat is the flexible path. You describe a requirement and LoopForge opens a conversation workspace. It can use project roles, work in a Docker sandbox, interact with a repository and prepare a delivery artifact.

Standard Operating Procedures (SOPs) are the repeatable path. An SOP is a graph of steps such as:

  • Agent: clarify the requirement or implement a change.
  • Script: run deterministic commands such as tests and linters.
  • Human gate: pause for a person to inspect and approve.
  • Sub-SOP: call a reusable workflow.

This separation matters. A successful test exit code is evidence that a command completed; an agent saying "I tested it" is not the same evidence. A human gate also turns approval into an explicit workflow state.

LoopForge conversation workspace from the official repository

Official workspace visual. The README links to the product tour.

What is new in v0.4.0?

The November 17 release makes the delivery workflow more complete:

  • Scheduled tasks keep execution history and separate conversations.
  • Conversation workspaces can run native agent CLIs on bound execution machines.
  • Git repositories can use SSH with per-host personal keys.
  • Built-in SOPs gain a merge-request node that commits, pushes and opens a PR/MR.
  • The assistant asks before external MCP write operations.

Read the v0.4.0 release notes and changelog before upgrading. The release notes call out migration steps that may need to be applied manually.

Start a local instance

The documented installation path is:

git clone --branch v0.4.0 https://github.com/Tencent/LoopForge.git
cd LoopForge
./start.sh
Enter fullscreen mode Exit fullscreen mode

You need Git, curl, Python 3 and Docker with Compose. The release pins the platform image to ghcr.io/tencent/loopforge:0.4.0.

After startup, open the demo project and try the Platform check Β· script & approval SOP described in the README. It runs a script, produces a report and waits for human approval. This is a useful first test because it demonstrates the workflow engine without requiring a model API key.

For agent work, configure the executor credentials and connect a repository. LoopForge supports CodeBuddy, Codex and Claude Code executors. You bring your own provider credentials and pay model usage separately.

A practical team workflow

For a routine backend fix, create an SOP that clarifies acceptance criteria, creates an isolated workspace, lets an implementation role make the change, runs tests through a script node, asks a separate reviewer to inspect the diff and stops at a human gate. The final node opens the PR/MR and attaches run artifacts.

If a test command is wrong, fix the SOP once. If a reviewer rejects the patch, send it back with specific feedback. This is more useful than repeatedly asking an agent to remember the team's process.

Good fits include routine bug fixes, dependency updates, release checklists, documentation changes and repository onboarding. For a one-off experiment, using an agent CLI directly is simpler.

Security and operational trade-offs

Read this before exposing a server.

The default Compose setup starts Docker sandboxes through the host Docker socket. Access to that socket is effectively access to the host Docker daemon. LoopForge positions this configuration for trusted teams, not untrusted multi-tenant isolation.

The README also says new installations listen on all interfaces, sign-ups are enabled by default and signed-in users can read project data across the instance. Set AGENTFLOW_ALLOW_SIGNUP=false if registration should be closed, configure HTTPS before remote access, restrict network reachability and use least-privilege credentials. A login screen is not a tenant-isolation boundary.

The chat workbench is documented as Chinese-first for now, which may matter for English-only teams. The project is MIT-licensed, while third-party components retain their own licenses; review THIRD_PARTY_NOTICES.

LoopForge vs other workflow tools

Tool Better fit
LoopForge Repeatable software-delivery stages, repository sandboxes and human gates
n8n / Dify / Flowise General integrations and business or AI workflows
Claude Code / Codex directly Solo work and one-off coding sessions
CI/CD pipelines Deterministic tests, security checks and deployments

LoopForge is not a replacement for CI. Keep tests and deployment policies deterministic. Use agents for work that benefits from interpretation, and scripts plus human approval for the parts that need reliable gates.

Sources and further reading

What part of your team's development process would you turn into an SOP first?


I write about software, open source and developer tooling.

Top comments (0)