Durabull: BullMQ & Redis Admin
Administer BullMQ queues and monitor Redis health from your assistant. Diagnose backlogs, inspect jobs and workers, review schedules and alerts, retry failed jobs, and pause or resume queues with a specific request.
Connect
- ChatGPT: add the custom MCP server
https://app.durabull.io/mcpin plugin settings and complete browser OAuth. Connecting the server supplies tools, but does not install the nine workflow skills. For those, add the portable package atplugins/durabullto a local marketplace, install it from the Plugins Directory, and start a new conversation with the plugin enabled. See OpenAI’s complete-plugin installation guide. Supported MCP Apps hosts show the interactive explorer. - Claude: add the same URL as a remote connector and complete OAuth. The connector supplies tools and MCP Apps; the plugin supplies skills on hosts that support plugin installation.
- Claude Code: from the repository root, run
claude --plugin-dir ./plugins/durabull. To install from the repository marketplace, use/plugin marketplace add durabullhq/durabullfollowed by/plugin install durabull@durabull. Text tools and skills work without an embedded UI. - Self-hosted: generate an independent package with
bun run mcp:plugin --endpoint https://queues.example.com/mcp --out /tmp/durabull-plugin. This updates both host configurations and every skill's MCP dependency. Connect to your own origin, not the Cloud URL.
You need a Durabull account with access to a Redis connection and its BullMQ queues. OAuth grants, organization permissions and service-account policy determine which data and operations are available. Never paste Redis credentials or access tokens into chat.
Start with a real task
| Skill | Example request | Result |
|---|---|---|
| Setup | “Connect Durabull and show my environments.” | Verified access and connection selection |
| Fleet health | “Which production queues need attention?” | Scoped inventory, backlog, workers and incidents |
| Queue triage | “Why is the receipt queue backing up?” | Evidence for the bottleneck and next step |
| Inspect job | “Find job 1042 and explain its failure.” | Exact job identity, redacted evidence and gaps |
| Schedules | “When does the daily digest run next?” | Cron/interval, timezone, limits and recent failures |
| Redis health | “Is Redis memory pressure affecting jobs?” | Capacity, history, thresholds and sample coverage |
| Alert triage | “Why was this incident's notification not delivered?” | Delivery attempts, rule state and incident evidence |
| Job recovery | “Retry this failed receipt job once.” | State-checked retry/promotion and observed result |
| Queue control | “Pause receipts during maintenance.” | Verified pause/resume state |
Alert triage also handles requested acknowledgement, unacknowledgement, resolution, snooze and unsnooze. The embedded app supports paginated inventories, exact job-ID search across queues, logs and stacktraces, recurring schedules, incident and rule inspection, Redis history, light/dark themes and host-supported fullscreen. “Ask to…” buttons send explicit requests to your assistant; data calls remain authorized on the server.
Boundaries
This is a BullMQ operations integration, not an arbitrary Redis command interface. It does not edit job payloads or Redis keys, delete jobs, purge queues, scale workers, create/edit schedules or configure alert rules. A retry being queued does not prove success. A snapshot is not continuous monitoring. Missing or stale samples remain visible as uncertainty.
Maintainers
plugin.json and mcp.json are canonical. bun run mcp:plugin generates Claude compatibility files and agents/openai.yaml for all skills. bun run mcp:check checks schemas, coverage, bundle freshness, protocol and policy. Run claude plugin validate plugins/durabull from the repository root.
evals.json contains 20 workflow scenarios and 8 boundary scenarios for real-host testing. Automated catalog checks verify that the scenarios reference real tools; they do not measure model routing or live OAuth behavior. Screenshots under assets/ use clearly labeled preview fixtures and contain no customer data.
The package is ready for local installation and submission preparation. Public directory availability requires each platform's publisher review; this repository is not evidence of approval. See the repository's docs/mcp-plugin-release.md for the submission and validation record.