LoopForge: self-hosted coding agent orchestration
Use LoopForge when a team needs repeatable coding-agent delivery with shared roles, review gates, and run history rather than isolated CLI sessions. It turns coding agents into governed multi-agent workflow engine participants, with Docker sandboxes, approvals, artifacts, and pull requests on your own infrastructure. Checked 2026-11-18: MIT license, last push 2026-11-18.
Workflow and delivery capabilities
- ✓Main Bot accepts work through chat, opens a Docker workspace, delegates to project roles, and can finish with a pull request or merge.
- ✓The visual SOP designer supports agent, script, human-gate, sub-SOP, and preview-build nodes, including rework paths when a step fails.
- ✓Agent roles keep prompts, executors, models, rules, Skills, MCP tools, artifacts, and gates separate across planning, implementation, review, and testing.
- ✓Docker sandboxes check out project repositories and run SOP nodes and conversation workspaces.
- ✓Runs retain node status, tool traces, logs, artifact snapshots, previews, and checkpoint history for inspection or resumption.
- ✓Requirements stay connected to work produced through chat or an SOP, including pull request status, merge detection, and cost reporting.
Where LoopForge fits
- •Run a bug-fix workflow with separate planning, implementation, review, testing, and human acceptance roles.
- •Deliver a backend or frontend project through repeatable, published SOP versions.
- •Generate product documents through agent steps without connecting an existing repository.
- •Open pull requests from repository changes while retaining approval and execution evidence.
- •Operate scheduled coding-agent tasks with workspace, SOP, and MCP access.
- •Try a script-and-approval workflow without model credentials through the built-in demo project.
Deploying the platform
Clone the repository and run the installer with Git and Docker already available: bash git clone https://github.com/Tencent/LoopForge.git cd LoopForge ./start.sh The installer checks Docker, Compose, Python 3, and curl. It pulls the release image pinned in `compose.yaml`, creates MySQL and platform services, generates configuration and session secrets, and installs built-in sandboxes, starter SOPs, and a welcome project. Host Go and Node.js are not required for the packaged deployment. Use `./start.sh dev` to build from source, or `./start.sh logs`, `stop`, `restart`, and `pull` for management. GitHub authorization and model API keys are configured after startup.
Chat and SOP workflows
In chat mode, Main Bot creates a conversation workspace, checks out repositories, and calls project roles as subagents. You can queue or steer messages, pause the response, switch executors, and inspect generated files. For structured delivery, publish or copy an SOP and run its nodes in dependency order. Each node can invoke an agent, run a deterministic script, wait for a person, call another SOP, or build a static preview. A failed or rejected step can be retried, rerun, or sent back upstream. Both modes retain artifacts, logs, requirement links, and delivery status.
Who should adopt LoopForge
LoopForge fits software teams that need more consistent agent-assisted delivery than a single coding CLI provides, especially where review ownership, approval, and audit evidence matter. It also suits platform teams willing to operate Docker-based infrastructure and configure model credentials, repository access, and agent CLIs. Individual developers who mainly want local coding-assistant workflows may prefer LoopForge Local or the CLI directly.
Strengths
- ✓Separates planning, implementation, review and testing into roles with independent prompts, executors, models, rules, Skills and MCP tools.
- ✓Combines Agent, Script, Human gate and Sub-SOP nodes in one visual workflow, with dependencies, approval pauses, retries and reruns from any node.
- ✓Runs coding-agent CLIs inside Docker sandboxes with checked-out repositories, checkpoints, logs, tool traces and per-node artifacts.
- ✓Supports both exploratory chat with Main Bot and repeatable SOP-based delivery, with requirements, PR/MR status, approvals and cost tracked across both paths.
- ✓Self-hosted under the MIT license, with GitHub, GitLab, Gitea, Gogs, Skills, MCP and local execution-machine integration documented.
Deployment and platform constraints
- △The README says the default sandbox model gives LoopForge access to the host Docker socket, which can control the host daemon. That setup is intended for trusted teams, not untrusted multi-tenant isolation.
- △New installations listen on all interfaces, while signed-in users can read project data throughout the instance. HTTPS, signup controls, and network restrictions are needed before wider exposure.
- △Executor setup is uneven: Claude Code and several other CLIs install on first use, Cursor CLI and Antigravity use provider installers and are not installed into SOP sandboxes, while ZCode must already exist on the service or execution-machine PATH.
- △Team members using the CLIs that retain per-run private home directories must supply their own model keys and Base URL settings. Shared administrator credentials are described for other executors, but credential ownership still varies by CLI.
- △The README identifies the chat workbench as Chinese-only for now, while management pages support English and Chinese. LoopForge Local can run built-in workflows, but exporting a platform SOP for standalone local execution is only planned.
- △The MIT license covers LoopForge, while third-party components retain separate licenses and notices. Operators still need to review those notices when assembling and distributing the service.
Comparing other workflow options
Frequently asked questions
LoopForge’s demo project does not require a model API key because its sample SOP contains only a script and a human approval gate. Chat conversations and agent-backed SOP steps require credentials for the selected executor.
GitHub is optional for LoopForge chat, document generation, and new projects because those tasks can run in the workspace. Connect a Git host when you need to modify an existing repository or open a pull request or merge request.
Use LoopForge chat while exploring a requirement or iterating on a fix. Choose an SOP when the work needs fixed roles, scripts, reviewers, approval gates, and repeatable delivery; Main Bot can also start and control SOP runs.
LoopForge can use Claude Code and Codex as executors. Each SOP role can select its executor independently, while credentials, models, sandboxes, workflow ordering, and delivery evidence remain managed by LoopForge.
LoopForge can be exposed, but the README treats a fresh installation as suitable for a trusted team. Add HTTPS, restrict access until an administrator exists, review the account model, and control signup before opening the service publicly.
LoopForge uses the MIT license. Third-party components remain under their own licenses and notices, which are recorded separately in the repository’s `THIRD_PARTY_NOTICES.md` file.
The problem it solves
Coding-agent CLIs can vary how they plan, test, review, and deliver the same request. LoopForge addresses that inconsistency by wrapping each agent step in a defined role, sandbox, workflow stage, and approval process, while linking every run back to its original requirement.
Related repositories
Is LoopForge worth your time?
ChatGPT, Claude and Perplexity can all read this page. Ask one of them what it makes of LoopForge.
