In part 6 we turned an AI assistant into a helper that knows your experience. But it is blind: it can't see recruiter emails, your interview calendar or your application tracker. You keep copying things by hand.
MCP fixes that. This part is for people willing to spend an hour on setup to save a few minutes on every small task afterwards. And for anyone who wants to understand the risks, because giving AI access to your mail is serious.
What MCP is, briefly
MCP (Model Context Protocol) is an open protocol that lets an AI app connect to external tools and data. Anthropic created it in November 2024 and in December 2025 donated it to the Agentic AI Foundation under the Linux Foundation. Today MCP is supported by Claude (the app and Claude Code), ChatGPT, Cursor, VS Code and others.
Think of it as USB for AI. There is a client (Claude Desktop, Claude Code) and a server (a small program that can, say, read Gmail or drive a browser). The server describes its tools, the model decides when to call them, and you approve the actions.
Servers come in two kinds:
- remote, reached by URL with an OAuth sign-in. Claude calls these connectors. The easiest to set up.
- local, running on your computer (usually via
npxoruvx). Needed when the server works with your files or browser.
Connectors that help a job search
| Task | Connector | Kind | Note |
|---|---|---|---|
| Recruiter emails | Gmail (built into Claude) | Remote, official | Search and read mail, drafts |
| Interviews | Google Calendar (built into Claude) | Remote, official | Check whether your connector can create events or only read them |
| Documents, CV | Google Drive (built into Claude) | Remote, official | Read documents |
| Tracker in Notion | Notion MCP | Remote, official | https://mcp.notion.com/mcp, OAuth sign-in |
| Tracker in Google Sheets | Google's official server (Preview) or community mcp-google-sheets | Remote or local | The official one needs your own OAuth client, more setup |
| Job search | Indeed MCP | Remote, official | Search, job details, company data. Geared to markets where Indeed operates |
| Company career pages | Playwright MCP (Microsoft) | Local | AI drives a browser: open a page, read it, take a screenshot |
| Same, via Chrome | Chrome DevTools MCP (Google) | Local | Aimed at development, but reads pages too |
| Company research | Exa, Firecrawl, Brave Search | Remote or local | Web search and page reading, API key required (Exa has a rate-limited anonymous mode) |
| Local files | Filesystem (MCP reference server) | Local | Access only to the folder you name, e.g. your CV versions |
| Calls | Cal.com, Calendly | Remote | Both services have official MCP servers |
To be honest about the Ukrainian market: we found no official MCP servers for Djinni, DOU, Work.ua or robota.ua. For them you still have RSS feeds (see part 2) and the browser.
Setting up Claude Desktop
Remote connector (Notion, Indeed, your own URL):
- Open settings, the Connectors section.
- For built-in ones (Gmail, Calendar, Drive), click Connect and sign in to Google.
- For others, click Add custom connector, enter a name and the server URL, and complete the sign-in.
- In a chat, enable the connectors you need via the "+" menu → Connectors.
The free plan allows one custom connector; paid plans allow more. A remote server must be reachable from the internet, because Claude connects to it from its cloud.
Local server (Playwright, Filesystem): either install it as a Desktop Extension (.mcpb file) in the Extensions section, or add it to claude_desktop_config.json:
Restart the app after editing the config. Menu names change between versions, so if something doesn't match, look for Connectors or Extensions in settings.
Setting up Claude Code
If you're a developer and already use Claude Code in the terminal:
The /mcp command inside Claude Code shows server status and lets you sign in. A handy trick: make a job-search folder with a CSV tracker, your CV versions and interview notes. Claude Code works with files directly, so for a simple tracker you don't even need a Sheets server.
Three scenarios that really save time
1. Morning review.
Find emails from recruiters and companies in Gmail from the last 3 days. For each: company, role, what they ask for, whether I need to reply. Compare with my Notion tracker and update statuses. Draft replies but don't send them.
2. Interview prep.
Check my calendar for interviews this week. For each, find the posting and the thread in my mail, open the company's careers page and put together one page: the product, the stack, who I'm meeting, 10 likely technical questions for the stack.
Practise those technical questions out loud rather than with the AI: ShortLesson has spoken answers with grading for JavaScript, React, SQL, Git and HTTP.
3. Collecting openings from career pages.
Here is a list of 15 companies I'm interested in. Open their careers pages and find open frontend or full stack JavaScript roles. Add new ones to the tracker with a link and date. Don't apply to anything.
What NOT to automate
- Mass auto-apply. Technically, an agent with Playwright can fill in forms. But the answers will be generic, you don't control what is sent, and companies remember candidates in their ATS. The whole value of an application is that it's precise.
- Scraping LinkedIn. The LinkedIn User Agreement (section 8.2) explicitly bans bots and scraping. Community LinkedIn MCP servers run through your logged-in session, and the author of one of the most popular ones states in the README that it breaks LinkedIn's rules and can get your account banned. The hiQ v. LinkedIn case ended in 2022 with hiQ paying 500,000 dollars and having to delete the data it had collected.
- Job scrapers built on JobSpy and similar libraries pull data from LinkedIn, Indeed and Glassdoor. Same risk: platform rules and bans.
- Sending emails and messages without your click. Drafts yes. Sending by hand only.
Security: what to know
A mail connector means the text of every email lands in the model's context. And an email can contain instructions for the AI that you can't see. This is called prompt injection. In 2025 researchers at Invariant Labs also demonstrated "tool poisoning": a malicious MCP server hides instructions in its tool descriptions, which the model sees and you don't.
Practical rules:
- Connect only official servers or ones whose code you can check. Look at who maintains them.
- Don't enable every connector at once. A morning review needs Gmail and the tracker; the browser is unnecessary there.
- Don't combine "reads outside content" (mail, web) with "sends" without your approval.
- Keep per-action approval on, at least for writing and sending.
- Keep API keys and tokens in a password manager, not in notes. Some servers put the token right in the URL, so that URL is a secret too.
- Give Filesystem access to one folder only, not the whole disk.
Our take
MCP is useful not so AI can look for a job instead of you, but to remove the copying between tabs. Start with two connectors: mail and tracker. If you're using them daily after a week, add the calendar and the browser.
Next
The last one, part 8, covers interview prep: mock interviews, voice practice, coding platforms, and why AI hints during an interview are a bad idea.