Operate BullMQ
from your agent.
Durabull MCP connects Claude, ChatGPT, Codex and any MCP client to your queues. Ask what's failing, get the evidence, and approve the fix without leaving the chat.
https://app.durabull.io/mcpWorks where your team already works
- Claude
- ChatGPT
- Cowork
- Claude Code
- Codex
- Cursor
- Muse
- Grok
- GitHub Copilot
- Windsurf
- Gemini CLI
Illustrative conversation with fixture data. The Queue Explorer app renders inline in hosts that support MCP Apps; other clients get the same answers as text.
30
MCP tools
21 read · 9 write
9
OAuth scopes
writes never auto-granted
~20
App views
in one Queue Explorer
9
Agent skills
Claude Code · Codex · ChatGPT
Ask your agent
The whole queue lifecycle, in plain language.
Your agent works against live Durabull data with the access you grant it. Everything it does also shows up in the dashboard.
01Fleet health
Every queue, one question away.
One call checks up to 100 queues: backlog, failures, paused queues, queues waiting with no workers, open alerts, and the latest Redis sample. Your agent knows where to look before you do.
Anything in production I should worry about?
One thing: image:resize has 26 jobs waiting and no workers attached. Everything else is draining normally.
- Waiting · Active · Failed
- 154 · 12 · 7
- Waiting without workers
- image:resize
- Workers online
- 12
02Failure triage
Root cause in a sentence.
Find a job across queues by ID, read its logs and per-attempt stack traces, and get a deterministic failure explanation with a confidence level and a list of any evidence it could not see.
Job 48213 failed. Why?
It's in email:receipts. All 3 attempts hit the same SMTP rate limit. The stack trace and the failure reason agree, so confidence is high.
- Strongest signal
- 421 4.7.0 rate limited
- Attempts
- 3 of 3
- Confidence
- High
03Recovery
Fix it, with your say-so.
Retry failed jobs, promote delayed ones, and pause or resume queues. Writes need scopes you approve separately, check the job’s state first, and return a receipt showing the state before and after.
The provider is back. Retry 48213 and resume reports:nightly.
Done. Both changes went through, and I re-read them afterwards to confirm the new state.
- job 48213
- failed → waiting
- reports:nightly
- paused → running
Queued work is not proof of successful completion.
04Incidents & alerts
Run the incident from the chat.
See what’s firing by queue and rule, check whether each notification was delivered, and acknowledge, resolve, or snooze a noisy rule for up to seven days. Linked Linear issues close when an incident resolves.
Ack the receipt alerts and snooze that rule for an hour.
Both incidents are acknowledged under your name. “Receipt delivery failures” is snoozed until 15:04.
- Open incidents
- 2 · acknowledged
- Notifications
- email ✓ · webhook ✓ · linear ✓
- Snoozed until
- 15:04
Schedulers
Know what runs next.
Cron patterns, intervals, timezones, next runs, and recent failures for every scheduler on a connection.
What’s scheduled to run overnight?
- nightly-rollup
- 0 2 * * * · 02:00 ET
- reindex-hourly
- every 1h · 14:00
- digest-weekly
- Mon 09:00 UTC
Redis health
Catch memory pressure early.
Memory, CPU, clients, evictions, and fragmentation, plus up to 30 days of history checked against your thresholds.
Is Redis under memory pressure?
- Memory
- 64% · threshold 80%
- Evictions / min
- 0
- 6h peak
- 78% at 12:40
Not a wall of JSON.
A live dashboard in the chat.
When your host supports MCP Apps, Durabull answers with Queue Explorer: an interactive Durabull view inside the conversation. Click through it yourself, then ask your agent to act on what you find.
Try it · the real app, live
About 20 views, one app
Connections, queues, jobs, workers, schedules, incidents, alert rules, Redis health, and failure investigations. You can drill from fleet to stack trace without typing.
Changes go through your agent
The app can only call read tools. Buttons like “Ask to retry” send the request back to the conversation, so your agent and your host’s approval step stay in charge.
Your agent sees what you see
As you browse, the app tells the model which queue or job you have open, so “why is this one stuck?” just works.
Inline or fullscreen
Starts inline in the thread and expands to fullscreen when you need room. It follows your host’s theme and safe areas.
ChatGPT Work entrypoints
In ChatGPT Work, on supported surfaces, Queue Explorer can also open from the global sidebar and the conversation panel.
Holds no secrets
The app has no network access, no stored tokens, and no local storage. All data comes through authorized tool calls, redacted on the server.
Works with
The agents your team already uses.
Durabull MCP follows the open Model Context Protocol. Claude, ChatGPT, and Cowork get the interactive app, Claude Code and Codex get a plugin with skills, and any other MCP client can connect with the same URL.
Durabull MCP
OAuth 2.1 · 30 tools · MCP Apps
Documented setup
Claude
by Anthropic
Connector + interactive app
ChatGPT
by OpenAI
Plugin + interactive app
Cowork
by Anthropic
Connector + interactive app
Claude Code
by Anthropic
Plugin + 9 skills
Codex
by OpenAI
Plugin + 9 skills
Any remote MCP client · not yet tested by Durabull
Cursor
by Anysphere
Standard remote MCP
Muse
by Meta
Standard remote MCP
Grok
by xAI
Standard remote MCP
GitHub Copilot
by GitHub
Standard remote MCP
Windsurf
by Windsurf
Standard remote MCP
Gemini CLI
by Google
Standard remote MCP
Built-in workflows
4 prompts and 9 skills included.
Prompts give any client a one-click starting point. The plugin adds 9 skills for setup, fleet health, queue triage, job inspection, schedules, Redis health, alert triage, job recovery, and queue control.
/triage_failed_jobs
Group recent failures by root cause and recommend next steps.
/investigate_queue_backlog
Check workers, throughput, drain time, and Redis pressure.
/alert_activity_review
Open incidents, noisy rules, and delivery failures.
/connection_health_check
A healthy / degraded / unhealthy / unknown verdict, read-only.
Product names and logos are trademarks of their respective owners and are shown to indicate compatibility, not endorsement. Claude, Cowork, ChatGPT, Claude Code, and Codex have documented Durabull setups. Other clients are listed for standard remote MCP and have not been tested by Durabull; support varies by client, plan, and version.
Your rules, enforced on the server
Give your agent a key, not every key.
Read access is the default. Every write scope is opt-in on the consent screen, and each write tool checks the job or queue state before it changes anything.
Claude wants to access Durabull
Signed in as you · Acme Inc.
Read · default
- mcp:discoverServer info and the tool catalog
- mcp:jobs:readQueues, jobs, workers, schedulers
- mcp:logs:readJob logs and stack traces
- mcp:failures:readAlert rules and incidents
- mcp:diagnostics:readMetrics, Redis health, triage
Write · opt-in
- mcp:jobs:retryRetry failed jobs (can make changes)
- mcp:jobs:promotePromote delayed jobs (can make changes)
- mcp:queues:pausePause and resume queues (can make changes)
- mcp:failures:writeAcknowledge, resolve, snooze (can make changes)
OAuth 2.1, per person
Each teammate signs in with their own Durabull login, using PKCE and tokens bound to the resource. Agents only see the connections that person can already open.
Destructive actions can’t be granted
Delete, purge, obliterate, payload edits, and raw Redis commands have no scope, so no token can ever be granted them.
Redacted at the source
Secrets, connection URLs, bearer tokens, JWTs, and API keys are removed from every response before the model sees it.
Audit trail
Tool calls and resource reads are logged with the user, the tool, a hash of the input, and the result, including denials and rate limits. Logging is best-effort under heavy load.
Built for agent bursts
Separate rate budgets for reads, heavy diagnostics, and writes, per user and client, with Retry-After hints your agent can follow.
Treats your data as data
Server instructions tell the model to treat job names, payloads, logs, and alert text as untrusted input, never as instructions.
Connect
Connect in three steps.
You need a Durabull account. Each teammate connects their own agent with their own login.
Add the Durabull connector
In your agent’s settings, add a custom connector or remote MCP server with this URL.
https://app.durabull.io/mcpSign in and choose access
Your agent opens a Durabull sign-in page. Approve the read scopes, and any write scopes you want.
Ask for what you need
Try “Show my Durabull connections” or “What failed in production in the last hour?”
- 01Open Settings → Connectors and add a custom connector.
- 02Paste the Durabull MCP URL and sign in to Durabull.
- 03Queue Explorer renders inline in Claude and Cowork conversations.
https://app.durabull.io/mcp
Self-hosting? The same server runs at /mcp on your own Durabull URL, e.g. https://queues.example.com/mcp.
MCP, the Model Context Protocol, is an open standard that lets AI agents use tools in other software. Durabull MCP is our MCP server. It gives your agent 30 tools to inspect and operate BullMQ queues, jobs, workers, schedulers, alerts, and Redis health on your Durabull connections.
The future of ops is agentic
Bring Durabull into your agent.
- Triage from a chat
- Act with your approval
- Same login and access as the web app
https://app.durabull.io/mcp