An MCP server for Resque, mountable as a Rails engine. Lets MCP clients (e.g. Claude Code) inspect queues, workers, and failed jobs — and retry or clear failures — over a single authenticated Streamable HTTP endpoint.

• Ruby >= 3.2
• Rails >= 7.2
• Resque >= 2.7, < 4

Add the gem to your application's Gemfile:

# Gemfile
gem "resque-mcp"

Then run bundle install. Or install it yourself with gem install resque-mcp.

Mount the engine and configure the auth token:

# config/routes.rb
mount Resque::Mcp::Engine => "/resque-mcp"

# config/initializers/resque_mcp.rb
Resque::Mcp.configure do |c|
  # token could be created by e.g.: `bin/rails runner 'puts SecureRandom.base58(32)'`
  c.auth_token = Rails.application.credentials.dig(:resque_mcp, :token)

  # Optional: which job-args keys to mask as [FILTERED] in tool responses.
  # Defaults to your app's config.filter_parameters; an explicit list
  # replaces it (merge yourself if you want both):
  # c.filter_parameters = Rails.application.config.filter_parameters + [:iban]
end

The token is required — the endpoint answers 503 until one is configured, and 401 on any request without a matching Authorization: Bearer header. The engine talks to whatever Resque.redis your app already configured; it never opens its own Redis connection.

Job arguments shown by any tool are filtered through ActiveSupport::ParameterFilter before preview/truncation, using your Rails filter_parameters by default — the same keys you hide from your logs are hidden from the model. Filters match hash keys (at any depth, same semantics as Rails log filtering, including anchored dot-notation like /\Acredit_card\.code\z/); positional scalar args have no key and pass through. Set c.filter_parameters = [] to disable.

Scope honestly stated: filtering covers job args only. Exception messages and backtraces in failure records are shown unfiltered (Rails doesn't scrub those from logs either) — a secret interpolated into an exception message will be visible, so treat error text accordingly.

Connect Claude Code:

claude mcp add --transport http resque https://your-app.example.com/resque-mcp \
  --header "Authorization: Bearer <token>"

Then ask, e.g., "How is my Resque doing?"

The tool surface is read-only so far (failure retry/clear are planned). Every tool response returns structured JSON alongside a text body and ends in a meta footer naming the Rails environment and the Redis target (with any credentials stripped), so you always see what you are talking to.

Global Resque health snapshot; takes no parameters.

{
  "pending": 87, "processed": 1093421, "failed": 1342,
  "queues": 6, "workers": 12, "working": 3,
  "queue_sizes": { "default": 3, "mailers": 0, "imports": 84 },
  "meta": { "environment": "production", "redis": "redis://prod-redis:6379/0/resque" }
}

• pending and queue_sizes come from a single snapshot, so pending always equals the sum of queue_sizes.
• processed is Resque's lifetime counter; failed is the current number of records in the failed list.
• workers counts registered workers, working those currently running a job.

List all queues with sizes, or inspect a single queue. With queue and include_jobs: true it pages through the queue's pending job payloads:

{
  "queue": "imports", "size": 84,
  "jobs": [
    { "class": "ImportWorker", "args_preview": "[812, \"s3://bucket/batch-7.csv\"]" },
    { "class": "ImportWorker", "args_preview": "[813, \"s3://bucket/batch-8.csv\"]" }
  ],
  "page": { "total": 84, "offset": 0, "limit": 2, "returned": 2, "has_more": true, "next_offset": 2 },
  "meta": { "…": "…" }
}

Every paginated response carries this page envelope. Limits are capped at 100 — a larger request is clamped and the clamp noted in the response.

List registered workers with their current state, optionally filtered by state (working / idle / all):

{
  "workers": [
    {
      "id": "worker-host-1:4021:imports,default", "state": "working",
      "queues": ["imports", "default"], "started": "2026-06-30 04:12:09 UTC",
      "processed": 48211, "failed": 12, "heartbeat_expired": false,
      "current_job": { "queue": "imports", "class": "ImportWorker",
                       "args_preview": "[812, …]", "run_at": "2026-07-02T09:14:55Z" }
    }
  ],
  "counts": { "total": 12, "working": 3, "idle": 9, "heartbeat_expired": 1 },
  "page": { "total": 12, "offset": 0, "limit": 50, "returned": 12, "has_more": false, "next_offset": null },
  "meta": { "…": "…" }
}

• heartbeat_expired: true flags a worker whose last heartbeat is older than Resque.prune_interval — it is likely dead and its record stale.
• counts always covers all workers; a state filter only narrows the workers list (and its page envelope).

Page through failed jobs, newest first, as compact records (truncated error, args preview, no backtrace). Optionally filter by class_name.

{
  "failures": [
    {
      "index": 1341, "failed_at": "2026/07/02 08:59:12 UTC",
      "queue": "imports", "class": "ImportWorker",
      "args_preview": "[812, \"s3://bucket/batch-7.csv\"]",
      "exception": "PG::ConnectionBad",
      "error": "could not connect to server: Connection refused… (truncated)",
      "retried_at": null
    }
  ],
  "page": { "total": 1342, "offset": 0, "limit": 20, "returned": 20, "has_more": true, "next_offset": 20 },
  "meta": { "…": "…" }
}

• A failure's index is its position in the failed list, deliberately not called an id: indexes shift when failures are removed, so re-list after any removal.
• Under a class_name filter, next_offset is a raw scan cursor — continue paging with the returned value, never compute offset + limit yourself. Filtered totals cost a full-list scan and are marked "total_note": "scan".

Full detail for one failed job by index: complete args, exception, full error, backtrace (capped at 30 frames, with backtrace_omitted), and the worker that failed it.

All truncation anywhere in the tool surface is marked in-band ("… (truncated)", backtrace_omitted), so a model always knows when it is seeing an excerpt.

Both failure tools support Resque's redis_multi_queue failure backend: pass the failure-queue name as queue. On that backend the tools require it and answer with the list of failure queues when it is missing. On the default backend, queue can simply be omitted.

After checking out the repo, run bin/setup to install dependencies. Then, run bundle exec rake to run the tests and standardrb. You can also run bin/console for an interactive prompt.

To release a new version, update the changelog and the version number in version.rb, commit it, and then run bundle exec rake release, which will create a git tag for the version, push git commits and the created tag, and push the .gem file to rubygems.org.

Bug reports and pull requests are welcome on GitHub at https://github.com/jbockler/resque-mcp. This project is intended to be a safe, welcoming space for collaboration, and contributors are expected to adhere to the code of conduct.

The gem is available as open source under the terms of the MIT License.