The 50 MCP Tools Every Production AI Agent Actually Needs

A production AI agent is roughly 90% tools. The model does the reasoning; the tools do the work — searching, reading files, querying databases, sending email, opening pull requests. Building each tool as an MCP (Model Context Protocol) server means it's plug-and-play: standardized interface, portable across agents and SDKs, and testable in isolation. Here are the 50 tools that cover the vast majority of real production use cases, grouped by the kind of agent that needs them.

Core / shared tools (used by almost every agent)

These 12 are the baseline. Almost any agent needs some subset:

http_request · web_search · web_scrape · read_file · write_file · list_dir · sql_query · vector_search · send_email · slack_post · schedule_task · get_datetime

Support-agent tools

A support agent lives or dies on its ability to look things up and act on tickets safely:

kb_search · kb_ingest · ticket_lookup · ticket_create · ticket_update · ticket_close · escalate_to_human · customer_lookup · order_lookup · refund_request (HITL-gated) · canned_response_fetch · sentiment_score · csat_log

Note refund_request is gated behind human-in-the-loop approval — a tool that moves money should never fire autonomously.

Research-agent tools

A research agent needs to gather from many sources, then synthesize and cite:

search_news · search_academic · search_web · fetch_url · extract_article · summarize_doc · cite_source · compare_sources · export_markdown · export_pdf · dedupe_findings · fact_check · translate

Code-agent tools

A coding agent needs to read a repo, make changes, and prove they work — with the dangerous actions gated:

repo_clone · repo_read · code_search · run_tests · lint · git_branch · git_commit · open_pr (HITL-gated) · read_issue · apply_patch · run_shell (sandboxed, HITL-gated) · dependency_lookup

The pattern that makes the other 950 tools cheap

Fifty is a starting set, not a ceiling. The real leverage is a single wrapper pattern: once you've built one MCP tool with a clean interface and a test harness, every additional tool is a thin wrapper around an API call plus a test — stamped out, not engineered. That's why "50 tools" is a reasonable v1 and "add your own in one line" is a reasonable promise. The economics of tools are: the first is expensive, the rest are margin.

Why MCP specifically

MCP is becoming the standard interface between agents and tools. Building on it means your tools aren't locked to one framework or one vendor — the same tool works whether the agent runs on LangGraph, CrewAI, or a vanilla loop, and third-party MCP servers drop in without glue code. It's the USB-C moment for agent tooling, and betting against standards here would be a mistake.

All 50 of these ship with agentFast as MCP servers, each with a passing test, grouped by the production-layer agent that uses them.