MCP is the plug that lets Claude Code read your tickets, query your logs, and drive your scanners. Here is how to wire it up fast without handing over the keys to the kingdom.
See the workflow in action, tap through the tabs below:
// add a remote server, then check it
claude mcp add --transport http sentry https://mcp.sentry.dev/mcp claude mcp list # inside a session, finish OAuth with: /mcp
// project scope writes .mcp.json you can commit
claude mcp add --scope project --transport http sentry https://mcp.sentry.dev/mcp claude mcp add --scope user --transport stdio airtable --env AIRTABLE_API_KEY=YOUR_KEY -- npx -y airtable-mcp-server
// a read-only incident triage prompt
Using only read tools from the sentry server, list the 5 most frequent unresolved errors from the last 24 hours. For each, give the release that introduced it, the likely owner, and one suspected root cause. Do not resolve, assign, or comment on anything.
Treat every server as untrusted input. A connected server can feed Claude text that tries to steer it (prompt injection). Pin to servers you trust, prefer read-only scopes, and review the approval prompt for project .mcp.json files before you click yes. Outputs over 25,000 tokens get cut off by default, so narrow your queries.
What MCP Actually Is
Model Context Protocol (MCP) is an open standard that lets Claude Code talk to outside tools through small servers. Point it at an issue tracker, an error monitor, a database, or a scanner, and Claude can read real data and take real actions instead of guessing from your paste buffer. If you already use Claude Code for DevOps and security workflows, MCP is the next upgrade: less copy and paste, more work done where the data lives.
Quick Setup
Remote servers over HTTP are the recommended transport. One command adds one, and claude mcp list shows status. For OAuth servers, run /mcp inside a session and follow the browser login. Local servers run as processes, and you must put -- between Claude's flags and the server command.
claude mcp add --transport http notion https://mcp.notion.com/mcp claude mcp add --transport stdio airtable --env AIRTABLE_API_KEY=YOUR_KEY -- npx -y airtable-mcp-server
Pick a scope on purpose. Local (the default) is just you in this project. Project writes a .mcp.json you commit so the team shares it. User follows you into every project. Sharing config in the repo is the big win for teams, and Claude asks for approval before using project servers in an interactive session.
Rule of thumb: start read-only, in local scope, on one server. Promote it to project scope only after it has earned trust.
The Mindset: Connectors Are Coworkers With Root
A server is not a plugin you forget about. It is a colleague whose output flows straight into your agent's context. Give it the least access that gets the job done, and assume any text it returns could be hostile. The docs warn plainly that untrusted servers can expose you to prompt injection, so trust is the whole game. Build habits around it: scoped tokens, read-only first, human approval for anything that writes.
7 Workflows Worth Stealing
1. Incident triage from your error monitor
Connect your monitoring server and ask for the top unresolved errors, their releases, and likely owners. Use the triage prompt in the widget above. You get a ranked list in seconds instead of ten browser tabs.
2. Ticket to pull request
With an issue tracker server attached, try: "Read ticket ENG-142, summarize the acceptance criteria, then propose a plan before touching code." The plan-first phrasing keeps Claude from charging ahead.
3. Database reality checks
Point a read-only database server at a replica and ask: "How many rows in the audit table lack a created_by value? Show me the query you ran." Always ask for the query so you can verify it.
4. Security scanner triage
Attach a scanner or dependency-check server and prompt: "List findings above medium severity, group them by package, and tell me which are reachable from our code." Reachability analysis is where an agent beats a spreadsheet.
5. Docs and runbook lookup
Connect your wiki so Claude can cite the real runbook during an incident: "Find the failover runbook and check my current steps against it. Quote any step I skipped."
6. Cross-tool change summaries
Ask for a release note that pulls merged pull requests and linked tickets together: "Draft release notes for this week, grouped by service, flagging anything customer visible."
7. Compliance evidence collection
Use read-only servers to gather screenshots-free evidence like config values and access lists, then have Claude format them for your auditor. Humans still sign off, but the gathering drops from days to an hour. Want the full skill path? See our courses.
Prompt Patterns That Make Connectors Behave
Three habits separate people who get real value from people who get noise. Name the server in the prompt, for example "using only the sentry server," so Claude does not wander to a different tool. State the boundary, such as "read only, do not modify anything," and Claude will respect it far more reliably than a silent default. Finally, ask for receipts: the query it ran, the ticket IDs it read, the link it cited. Receipts turn an answer you have to trust into an answer you can check in ten seconds.
Also think in small loops. Ask one narrow question, read the result, then ask the next. A five step investigation as five short prompts beats one giant request that burns context and hides where it went wrong. When a result looks surprising, ask Claude to show the raw tool output before you act on its summary. That single step catches most misreadings early, and it is the cheapest safety net you will ever add to an agent workflow.
Try this today: run claude mcp list, remove anything you no longer use, and write one sentence per remaining server describing what it is allowed to touch.
Safety and Gotchas
First, prompt injection. A ticket, log line, or web page fetched through a server can contain instructions aimed at your agent. Keep write tools behind approval prompts, and never connect a server you have not vetted. Second, secrets. Do not paste tokens into shared config. Use headers with environment variable expansion in .mcp.json so credentials stay out of git. Third, remote servers that need auth show as "Needs authentication" until you run /mcp or claude mcp login. Fourth, a failed server usually means a bad URL, a missing -- separator, or an expired login, so check claude mcp get for the detail.
Usage and Cost Tips
MCP output costs context. Claude Code warns when a tool returns more than 10,000 tokens and caps output at 25,000 by default, and you can raise that with the MAX_MCP_OUTPUT_TOKENS environment variable. Do not raise it first. Ask narrower questions, filter by date or severity, and request summaries. Claude Code also uses tool search by default, so many connected servers do not flood every prompt with tool definitions. Still, disable servers you are not using with /mcp so the list stays short and the attack surface stays small.
FAQ
Is MCP the same as a Claude Code plugin or skill?
No. MCP connects Claude to external tools and data. Skills package instructions and know-how. They work well together: a skill can tell Claude how to triage, and an MCP server gives it the data to triage.
Should I use HTTP or stdio servers?
Use HTTP for hosted services, since it is the recommended transport and handles OAuth. Use stdio for local tools that run as a process on your machine. SSE is deprecated, so pick HTTP when a server offers both.
Is it safe to commit .mcp.json to my repo?
Yes, if it holds no secrets. Reference credentials through environment variables, and remember teammates must approve project servers before they run.
Your Move
Pick one read-only server this week, wire it in, and run the triage prompt against real data. Once that feels boring, add a second. Want guided practice building agent workflows for DevOps and security careers? Explore our courses.
