Tencent LoopForge: Self-Hosted SOPs That Orchestrate Claude Code, Codex and CodeBuddy

Tencent published LoopForge as a self-hosted control plane for coding-agent delivery — chat with Main Bot for exploratory work, or run SOPs (standard operating procedures) that chain agent roles, shell scripts, and human gates in Docker sandboxes. The repository trended with roughly 700+ stars and 174 stars gained in a day as of November 14, 2026, per GitHub counters.

LoopForge sits in the same problem space as Plannotator’s local review surface and Agent Autoloop’s background hub, but Tencent’s angle is team-scale delivery: published workflows, shared requirements, token-cost reports, and a web console — not only a single-developer CLI plugin.

TL;DR

Why a coding agent still needs a delivery layer

Asking an agent to “fix this bug” once is easy. Asking the same team to ship the same way every time — plan, implement, test, review, approve — is not. LoopForge’s README frames the gap explicitly: demos skip tests, agents review their own diffs, and nobody notices when a run stalls.

LoopForge adds roles (planner, implementer, reviewer, tester), scripts for deterministic steps, and human gates where a person signs off or sends work back to an earlier node. That matches how enterprises already run Claude for manufacturing plugins and AWS Bedrock agent routing, but here the orchestration is yours on your Docker host.

Two paths: chat with Main Bot or run an SOP

Both paths attach to the same requirements object, so whether code ships from chat or a SOP, history and cost roll up together — closer to ITSM than a one-off /loop in the terminal.

Human approval gates pause agent delivery until a teammate signs off

SOP nodes, per the README:

Failed checks can rework to earlier steps — the designer shows this with the built-in bug-fix SOP.

Install and first run

Tencent documents a Compose-first path that does not require Go or Node on the host:

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

./start.sh checks Docker, Compose, Python 3, and curl, pulls the release image pinned in compose.yaml, seeds built-in sandboxes and starter SOPs, and listens on 0.0.0.0:8090 by default. The first registered account becomes administrator.

Try without any model key: open the demo project, run the Platform check · script & approval SOP, approve the human gate, and watch a script-only path complete — useful for proving the engine before you paste API keys.

For air-gapped or slow ghcr.io regions, the README documents MIRROR=ghcr.nju.edu.cn ./start.sh to pull through a mirror and retag the official digest — verify with docker image inspect against the pinned digest in compose.yaml.

Developers hacking on LoopForge itself use bash scripts/dev.sh with Go 1.26.11+, per the repository docs.

Self-hosted Docker installs keep sandboxes and approval history on your hardware

Executors, GitHub, and sandboxes

Executors are the actual coding CLIs. CodeBuddy and Codex ship inside the image; Claude Code, OpenCode, Kimi Code, and Pi install on first use via npm and cache afterward. Cursor CLI and Antigravity use provider installers on chat first use but are not preinstalled in SOP sandboxes, the README warns — pick executors your sandbox can actually run.

GitHub is optional: generic product SOPs can deliver documents without a repo; bug-fix and backend SOPs need source. OAuth device flow is built in; SSH keys work for GitLab, Gitea, or Gogs with manual PR tooling.

Sandboxes talk to the host Docker socket — powerful and appropriate for trusted teams, not multi-tenant hostile users. That is the same trust model as Chrome DevTools MCP driving a real browser: convenience trades off against isolation.

LoopForge Local for single developers

Teams that want SOP ideas inside an existing CLI can install the portable edition:

npx -y loopforge-cli@latest install claude

Then invoke /start-devflow "<your request>" in Claude Code. Source lives under local/ in the main repository. Platform SOP export to standalone local runs is planned, not shipped — check the local guide before you depend on it.

How this compares to other orchestration stacks

LoopForge is closer to CI plus agent roles than to a chat UI skin. If you already standardized on Addy Osmani’s agent-skills packs, LoopForge gives those skills stage boundaries and approval cards in a shared inbox.

Token economics and observability

The README advertises token and cost accounting for SOPs and conversations at public list prices, plus engineering metrics exports. That matters when coding-agent bills surprise finance — LoopForge makes spend visible per requirement, not only per API key.

Runs expose node status, tool traces, worker logs, and per-node artifacts. When an agent step fails, you can retry or rerun from any node instead of replaying an entire chat from scratch — the operational difference between a toy demo and something a release manager will trust.

Security and governance notes

Report vulnerabilities through Tencent’s published security policy on the repository. For production:

• Treat Docker socket access as root-equivalent on the host.
• Restrict signup with AGENTFLOW_ALLOW_SIGNUP=false when you do not want open registration.
• Put HTTPS in front before exposing beyond a VPN.

LoopForge adds process around agents that already pass your CLI policies; it does not replace dangerous-tool guardrails inside each executor. Pair organizational SOPs with tool restrictions the way shadcn’s agent skill audit guidance recommends: expensive execute, cheap review.

HTTP API, MCP, and scheduled agent tasks

Beyond the browser console, LoopForge documents an OpenAPI-described HTTP API and MCP exposure so external systems can open requirements, enqueue SOP runs, or poll gate status. That matters when your incident bot already lives in Slack: the approval card can be a webhook into LoopForge’s Needs review inbox instead of a second chat thread nobody audits.

AI scheduled tasks reuse the same workspace machinery: recurring agent jobs with parameterized templates, versioned like SOPs, with full MCP and repository access when credentials allow. For teams comparing Cursor’s proactive Projects coordinator, LoopForge is heavier ops but offers shared history — every teammate sees the same run graph.

Built-in SOPs and the validation record

Tencent ships read-only blueprint SOPs for backend, frontend, bug-fix, and product flows. Copy one into your project before you edit nodes — the README warns that built-ins are templates, not drop-in production paths. The repository links a validation record and acceptance matrix describing a real four-role CodeBuddy run (plan, implement, review, summarize) with three passing tests and a human gate — useful as a realism check, not a guarantee your repository will pass unchanged.

If you maintain internal “definition of done” docs, map each DoD bullet to a Script node (lint, unit tests, SBOM) so agents cannot skip checks by “forgetting” a prompt line.

Cost reporting and model selection per role

LoopForge’s insights pages aggregate token usage at public list prices per requirement and per SOP version. That is not the same as invoice truth — your enterprise discount may differ — but it beats flying blind when Gen Z stacks AI subscriptions beside corporate API keys.

Per-role executor choice is deliberate: a planner might run an expensive model while the implementer uses Codex on a smaller tier. The README notes that requirement clarification, notification summaries, and scheduled assistant tasks currently pin to CodeBuddy-only model pickers on the server — mixing CLIs on those subsystems is rejected rather than silently swapped. Read that constraint before you standardize on Claude Code everywhere.

Ops checklist before you trust production delivery

When LoopForge is worth the ops cost

Choose LoopForge if multiple developers share coding agents, you need auditable delivery (gates, artifacts, rerun history), and you can host Docker on a trusted machine.

Skip it if one developer already lives entirely inside Claude Code with personal git hooks — Worktrunk or Plannotator may be lighter.

Watch the mirror docs if you deploy in regions where GHCR pulls stall; digest pinning is only meaningful if the pull actually completes.

Related reading

• Plannotator: annotate plans before agents run
• Agent Autoloop: background Claude Code and Codex hub
• Claude Code Projects proactive coordinator
• Tencent HY4 vision and terminal benchmarks
• What is the A2A protocol
• AWS Bedrock coding-agent routing patterns
• Track and cap cloud agent spend

Sources: Tencent LoopForge repository, LoopForge product site