Real-world MCP servers: file system, git, browser, databases

Three ways to see it

1. 1

   File system server. The official 'filesystem' server exposes a controlled folder to the model. Tools include list_directory, read_file, write_file, search_files. The server is scoped at launch to one folder; the model cannot escape upwards. A Pakistani example: scope it to /Users/analyst/fbr-returns where the team drops every PDF return for the day. The model can now read every PDF, summarise irregular ones, and write a report file back into the same folder. The server does not need to understand PDFs. The model does, because LLMs are general parsers. The combination is what produces value.

2. 2

   Git server. The 'git' server exposes a local repository: git_log, git_diff, git_status, git_show. The model can review recent changes, summarise commits in plain language, and find which engineer touched which file. Useful in a Pakistani software house running a fixed-price project for a government department: every Friday, the senior engineer asks Claude for a one-page report of all changes that touched the FBR integration this week. The model reads the git server, the file system server, and produces the report. What used to be two hours of manual review becomes five minutes of conversation.

3. 3

   Browser server. Playwright-based MCP servers expose web pages to the model: navigate, click, type, screenshot, read DOM. The Pakistani use case is the most exciting. SECP eServices has no first-class REST API. A browser MCP server lets the model log in, search a company, click through the corporate filings, and return the data. This is fragile (the page can change), it is legal only in scopes that respect terms of service, and it should always be rate-limited politely. But it is a bridge from the world of clunky portals to the world of structured AI workflows. Every Pakistani team that wishes a public portal had an API can build that bridge themselves.

The why-tree

Try this with Claude

AI-edge prompt to try with Claude or ChatGPT: 'I run operations at a 50-person Karachi fintech. Recommend a six-server MCP catalogue: file system, git, database, browser, Slack, and one Pakistani-specific server. For each, give the scope, the risk, and the one mitigation that brings it inside enterprise-grade governance. Output: a tight table.' Compare to your own catalogue.