Skip to main content
Glama
mcp-ops
by 2moulin

mcp-ops

Give your AI agent read-only eyes on your whole production stack.

license: MIT  node  TypeScript  Model Context Protocol  read-only by construction  services  tests

Stripe · Postgres · Neon · Vercel · Netlify · Render · Cloudflare · Resend · Postmark · Twilio · Sentry · Better Stack · Shopify · GitHub

One MCP server. Every service behind two tools that see across all of them. Nothing it can touch is ever written.


Stop opening a dozen tabs when prod breaks. Ask once:

> what happened in the last 90 minutes?

13:02  github   commit          fix: transfer funded in the settled currency
13:04  vercel   deploy ready    production win-beta [main]
13:11  stripe   payment         pi_3Q… 135.00 CAD
13:11  stripe   transfer.created tr_1Q… 128.00 CAD → acct_1N… (webhook pending)
13:11  resend   email delivered Your order is confirmed → j***@example.com
13:12  sentry   issue error     WIN-42 cannot read 'currency' (x3, first 2m ago)
13:14  uptime   incident        WIN HTTP 502

Two tools that look everywhere

ops_status

The whole stack in one screen. Start here.

ops_timeline

Deploys, commits, payments, errors, emails, incidents — merged and sorted. "What changed before it broke."

Plus per-service tools, and diagnostics that explain why: a blocked payout, a failed transfer, a broken build, a weak TLS setting.

Related MCP server: adf-mcp-server

Install

claude mcp add ops \
  -e STRIPE_SECRET_KEY=rk_test_... \
  -e DATABASE_URL=postgres://... \
  -e GITHUB_TOKEN=ghp_... -e GITHUB_REPO=owner/name \
  -- npx -y mcp-ops
{
  "mcpServers": {
    "ops": {
      "command": "npx",
      "args": ["-y", "mcp-ops"],
      "env": {
        "STRIPE_SECRET_KEY": "rk_test_...",
        "DATABASE_URL": "postgres://...",
        "GITHUB_TOKEN": "ghp_...",
        "GITHUB_REPO": "owner/name"
      }
    }
  }
}

Set only the services you use — one is enough. A misconfigured service is skipped, never the server.

Services & keys

Service

Variables

Key

Stripe Connect

STRIPE_SECRET_KEY

restricted, read scopes

Postgres

DATABASE_URL

read-only role or replica

Neon

NEON_API_KEY [NEON_PROJECT_ID]

API key

Vercel

VERCEL_TOKEN [VERCEL_TEAM_ID] [VERCEL_PROJECT]

team-scoped

Netlify

NETLIFY_TOKEN [NETLIFY_SITE]

personal token

Render

RENDER_API_KEY

API key

Cloudflare

CLOUDFLARE_API_TOKEN [CLOUDFLARE_ZONE]

Zone:Read, DNS:Read

Resend

RESEND_API_KEY

read key

Postmark

POSTMARK_SERVER_TOKEN

server token

Twilio

TWILIO_ACCOUNT_SID + TWILIO_AUTH_TOKEN

read scope

Sentry

SENTRY_TOKEN + SENTRY_ORG + SENTRY_PROJECT [SENTRY_URL]

project:read, event:read

Better Stack

BETTERSTACK_TOKEN

team token

Shopify

SHOPIFY_SHOP + SHOPIFY_ACCESS_TOKEN

read_orders

GitHub

GITHUB_TOKEN + GITHUB_REPO

read Contents, Actions, PRs

Read-only by construction

  • No tool writes. Postgres runs inside BEGIN READ ONLY with a timeout; anything but a single SELECT/WITH/EXPLAIN is rejected before it leaves the process.

  • Live Stripe keys refused unless STRIPE_MCP_ALLOW_LIVE=1.

  • PII masked everywhere: emails, phones, bank & tracking numbers to last 4. SMS bodies and env values never returned.

  • Secrets stay put. Error messages strip query strings. TLS verified by default. Every call times out, every output is capped.

  • Returned content (commit messages, subjects, order notes) is labelled untrusted so the agent treats it as data, not instructions.

Ask it

  • "ops_status, then tell me what needs attention."

  • "acct_1XYZ says they were never paid — diagnose it."

  • "Is our Cloudflare zone actually encrypting to origin?"

  • "Which Sentry issues appeared after commit a1b2c3d?"

  • "How many orders are stuck in pending, and since when?"

Develop

npm install
npm test        # every service vs a fake API + a real MCP client
npm run dev     # stdio server

A new service is one file (Provider: register, status, optional timeline) and one line in src/index.ts.

License

MIT © David Demoulin

Available Tools

13 tools
ops_statusStatus of the whole stackA

One line per connected service (stripe): balances, blocked accounts, failing webhooks, last deploy, open errors, bounced emails, database size and connections. Start here.

Inputs: no named parameters.

ops_timelineWhat happened, across every serviceA

One chronological list mixing deploys, commits, CI runs, payments, errors, emails, SMS, incidents and database operations from every connected service, for the last N minutes. The fastest way to answer "what changed before things broke".

Inputs: limit, sources, since_minutes.

stripe_balanceStripe balanceA

Available and pending balance per currency, for the platform (no account_id) or for one connected account.

Inputs: account_id.

stripe_diagnose_accountDiagnose a connected accountA

Explain why a connected account cannot charge or get paid out: disabled reasons, past-due requirements, capabilities, missing bank account, TOS.

Inputs: account_id (required).

stripe_diagnose_transferDiagnose a transferA

Given a transfer id (tr_...) or a charge id (ch_...), check what makes Connect transfers fail: settlement currency of the source charge vs the transfer currency, amount above the charge net, platform balance too low, destination with payouts disabled, refunds and disputes.

Inputs: id (required).

stripe_get_accountGet a connected accountA

Full status of one connected account: capabilities, requirements, external accounts (last4 only), dashboard type.

Inputs: account_id (required).

stripe_get_eventGet a Stripe eventA

One event with its payload (clipped), pending webhook count and the request that caused it.

Inputs: event_id (required).

stripe_list_accountsList connected accountsB

List the connected accounts of this Stripe Connect platform with their status (charges, payouts, requirements due).

Inputs: limit.

stripe_list_chargesList chargesB

Recent charges on the platform, or on one connected account (direct charges) when account_id is given. Shows destination, application fee and attached transfer.

Inputs: limit, account_id.

stripe_list_eventsList Stripe eventsA

Recent Stripe events. Filter by type ("payment_intent.succeeded") or by prefix ("payment_intent."). only_undelivered=true lists events whose webhook delivery is still failing.

Inputs: type, limit, only_undelivered.

stripe_list_payoutsList payouts of a connected accountB

Payouts from a connected account to its bank, with status, arrival date and failure reason.

Inputs: account_id (required), limit.

stripe_list_transfersList transfersC

Transfers from the platform to connected accounts: amount, currency, destination, funding charge, reversals.

Inputs: limit, since_days, destination.

stripe_list_webhook_endpointsList webhook endpointsC

Webhook endpoints of the platform with their status, API version and enabled events.

Inputs: no named parameters.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.4/5.0

    Scored across 13 tools

    Disambiguation5/5

    Each tool targets a distinct resource or action: list/get/diagnose split account lookup from troubleshooting, and transfer diagnostics are clearly separated from transfer listing. Even the two ops-level tools differ in purpose: status is a snapshot summary while timeline is a chronological event view.

    Naming Consistency4/5

    The stripe_ prefix with list_/get_/diagnose_ verbs is highly consistent, and ops_ follows the same resource_scope pattern. Minor deviations are stripe_balance (a noun instead of get_balance) and ops_status/ops_timeline being noun-only, but the pattern remains predictable.

    Tool Count5/5

    Thirteen tools cover the Stripe Connect operational surface at a reasonable granularity without redundancy. The count sits comfortably below the heavy threshold and each tool has a clear role, from account diagnostics to event inspection and cross-service status/timeline.

    Completeness4/5

    The surface covers the core read/diagnose workflows: accounts, balances, charges, transfers, payouts, events, webhooks, and aggregate ops views. Minor gaps remain, such as no dedicated refund/dispute list or single-transfer getter, but most can be worked around via events and diagnostic tools.

    Maintenance

    ActivitySlowing
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      B
      maintenance
      Enables read-only Kubernetes incident investigation through MCP tools for listing pods, describing resources, fetching logs, and searching runbooks.
      1
      -
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables read-only investigation of service incidents by retrieving metrics, logs, and deployment history, and correlating evidence for root-cause analysis.
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Provides read-only production observability through a small set of tools for health, job status, latest data, approved logs, and deployment metadata, enabling AI planning and review without a production write path.
      MIT