Execution modes
The execution mode decides where the analysis runs and who pays the model. Approval, tags and the messages to the customer always go through Loopback.
Managed
Each job runs in a private container in our cloud, created for the job and deleted at the end, with CPU, memory and time limits and network access only to GitHub, the model gateway and package registries. The repository is cloned with a short-lived token, and the models are paid by Loopback: the plan includes the fixes. The agent gets no provider key; it calls our model gateway with a token valid for its job.
Your runner
Jobs run in your GitHub Actions through the GitHub Action, or in your own container, see self-hosted runner. The model is paid by you, with your API key, Bedrock, Vertex, or a ChatGPT plan for Codex. Jobs do not run in our cloud: Loopback receives the diagnoses, the reply drafts and the reports, as described in security and data handling.
In your agent (MCP)
Your agent takes tickets from Loopback through the Model Context Protocol while you work, inside your IDE or CLI, with the subscription or key you already pay for. Jobs wait in the queue when nobody is working. See MCP setup.
Which mode for your answer
The setup asks one question, which agent you use and how you pay for it, and proposes the mode.
| Answer | Proposed | Why |
|---|---|---|
| I do not use agents, I want it to work | ManagedBuilt-in agent · API key | We run everything in our cloud, with tokens included in the plan. |
| I only have an API key | RunnerBuilt-in agent · API key | Our builtin agent runs in your GitHub runner with your key. |
| Claude Code with an API key, Bedrock or Vertex | RunnerClaude Code · API key | Claude Code runs in your GitHub runner with your key or cloud provider. |
| Claude Code with a Pro or Max subscription | MCPClaude Code · Subscription | A Claude subscription can only be used inside the official client: Loopback hands you tickets over MCP while you work. To work while you are offline, add an API key and use the runner. |
| Codex with a ChatGPT plan or a key | RunnerCodex · ChatGPT plan | Codex runs in your runner with your ChatGPT plan or your key. |
| Cursor, Windsurf, Zed or another IDE agent | MCPCursor · Subscription | Your agent pulls tickets from Loopback over MCP, inside your IDE. |
Compatibility matrix
Agents, how they run without a person at the keyboard, and the authentication allowed in each mode. Checked version by version.
| Agent | Runs without a person | Authentication in runner and Managed | Authentication in MCP mode | In Managed | Notes |
|---|---|---|---|---|---|
| Claude Code | claude -p | API key, Amazon Bedrock, Google Vertex AI | Subscription, API key, Amazon Bedrock, Google Vertex AI | Yes | Rule 6: no subscription OAuth outside MCP mode. |
| Codex | codex exec | API key, ChatGPT plan | Subscription, ChatGPT plan, API key | No | auth.json is a password; the refreshed file is persisted; one job at a time per credential. |
| opencode | opencode run | API key | API key, Subscription | No | Check flags and output per version. |
| Built-in agent | builtin | API key, Amazon Bedrock, Google Vertex AI | None | Yes | Default for Managed and for teams without an agent; in Managed it uses Loopback keys. |
| Cursor | No | None | Subscription, API key | No | MCP mode only. |
| Windsurf | No | None | Subscription, API key | No | MCP mode only. |
| Zed | No | None | Subscription, API key | No | MCP mode only. |
| Kiro | No | None | Subscription, API key | No | MCP mode only. |
| Antigravity | No | None | Subscription, API key | No | MCP mode only. |
| Continue | No | None | Subscription, API key | No | MCP mode only. |
| Cline | No | None | Subscription, API key | No | MCP mode only. |
| GitHub Copilot | No | None | Subscription, API key | No | MCP mode only. |
| Other MCP agents | No | None | Subscription, API key | No | MCP mode only. |