Copilot (experimental)

The Copilot is a chat assistant built into the dashboard. It reads your live queue state and can propose only three mutation types, Promote, Pause and Resume, which you confirm before they run. It is experimental, and stays off until you add a model.

Bring your own model

Open the Copilot (the button in the bottom right), then the settings gear, and pick a provider:

• Claude (Anthropic)
• ChatGPT (OpenAI)
• Gemini (Google)
• GLM (Z.ai)
• OpenRouter (one key, every model)
• Custom (any OpenAI-compatible endpoint: Groq, Together, Mistral, a local Ollama or LM Studio)

It is built on the Vercel AI SDK, so all of these work. Paste your own API key and set the model id. The key is kept in memory for the session only and is never written to disk; the provider and model are remembered. Named providers (including Z.ai and OpenRouter) always use the endpoint built into the dashboard: a stale or injected saved base URL is discarded. Only Custom can use an operator-supplied base URL, and it must be a non-empty http:// or https:// endpoint without credentials, a query, or a fragment.

Browser (CORS) note

The dashboard is a browser app, so the model call goes straight from your browser to the provider. Some providers allow that and some do not:

• Work browser-direct: Claude (with the safe browser header, added automatically), OpenRouter, Z.ai, and most OpenAI-compatible endpoints.
• Blocked by the provider: OpenAI and native Google Gemini block direct browser calls. Reach those through OpenRouter, or run behind a proxy.

What it can do

Read (runs immediately): list queues and counts, list and inspect jobs, DLQ stats and entries, server health and stats, workers, and crons.

Change (asks first): promote a delayed job, pause a queue or resume a queue. Those are the Copilot's only mutating tools. Cancel/remove, DLQ retry (single or bulk), completed-job requeue, generic retry and purge tools are intentionally absent.

DLQ retry is not made safe by a preliminary read: in Bunqueue v2.8.59 the later POST has no atomic generation/state/topology precondition and can hit a different job recreated under the same ID. Completed-job requeue is also unavailable because retryCompleted does not reconstruct dependency registration or original flow order.

Every mutating action shows a confirmation in the chat and only runs after you click Confirm. One chat turn is pinned to the server URL and Bearer token that were active when the turn started, so all of its reads stay on that same target. The confirmation label names that server. If the selected server or token has changed before approval, the write fails closed and no mutation request is sent.

Try it

In the live demo the Copilot answers against the demo fixtures, so with your own key you can watch it hold a real conversation and drive tools with no backend.