Best AI Agent Skills for Business Analysts (2026)

Business analysts spend most of their week moving information between formats. A query result becomes a pivot table, the pivot table becomes a chart, the chart becomes a slide, the slide becomes a decision, and somebody asks where the number came from. Agent skills are unusually well suited to that loop, because the loop is mechanical and the judgment sits at the edges.

The skills directory has around 1,300 entries and almost none of them are labelled for business analysts. The useful ones are filed under coding and data. Here's what earns a place in a BA's setup, grouped by which part of the week it fixes.

The shortlist

One note on the star column before anyone quotes it back at me. XLSX, DOCX, and PPTX all live in the anthropics/skills monorepo, which carries about 170,000 stars across every skill in it. Attributing that number to any single skill would be dishonest, so I've left it as "monorepo" and used install counts instead, which are per-skill and therefore meaningful.

Getting the data

MySQL MCP Server by designcomputer is the plainest win on this list if your warehouse speaks MySQL. It gives the agent a real connection, so instead of describing a query and pasting the result, you ask a question and the agent runs SQL against the database.

claude mcp add mysql-mcp-server -- npx -y designcomputer/mysql_mcp_server

Verified tier, community-reviewed, MIT, 1,360 stars, last updated 2026-08-02. That recency matters more than the star count for anything holding database credentials.

The obvious hazard: an agent with a live connection can write as well as read unless you constrain it. Point it at a read replica or a role with SELECT only. This isn't paranoia about the model, it's the same thing you'd do for any service account.

CSV Data Summarizer by coffeefuelbump covers the other common case, which is that somebody emailed you a file. It profiles the CSV, generates summary statistics, detects missing data, and produces quick visualizations through pandas. 450 stars, community tier, unreviewed, and last updated in October 2025. That date is the problem. Ten months without a commit on a skill that executes Python against your files is a reason to read the source before installing, not a reason to skip it, but read it.

Excel MCP Server by haris-musa handles the spreadsheets that never make it into a database. 4,113 stars, verified, MIT, last updated April 2026. It manipulates workbooks in place rather than reading them into context, which matters when the file is 40MB of quarterly actuals.

claude mcp add excel-mcp-server -- npx -y haris-musa/excel-mcp-server

Producing the artifact

This is where the official Anthropic skills carry the day, and it isn't close.

XLSX reads, writes, edits, and analyzes spreadsheets, including formulas, charts, and cell formatting, across .xlsx, .xlsm, .csv, and .tsv. Roughly 134,000 installs, audited, MIT.

claude skill add anthropics/skills/xlsx

DOCX does the same for Word: tables, headers, styles, images, complex formatting. About 115,000 installs. This is the one that turns a requirements conversation into an actual document with a version history instead of a wall of chat.

claude skill add anthropics/skills/docx

PPTX builds presentations with layouts, charts, speaker notes, and animations. About 96,000 installs. Steering committee decks are the single most automatable artifact in the BA job, and this does about 80% of one before you touch it.

claude skill add anthropics/skills/pptx

All three are audited and maintained by Anthropic, updated 2026-08-17. If you install nothing else from this article, install these.

MCP Server Chart by antvis adds 25+ visualization types generated from data rather than from a screenshot of data. 4,319 stars, verified, MIT. Its weakness is that it was last updated in May 2026, which is a longer gap than I'd like on an actively used dependency, though a chart renderer ages more gracefully than a database connector.

Requirements and specs

OpenSpec by Fission-AI is the surprise on this list. It's marketed at developers doing spec-driven development, but the underlying idea maps cleanly onto business analysis: write the specification first, keep it as the source of truth, and make every downstream artifact derive from it rather than drift from it. 65,203 stars, verified tier, MIT, updated 2026-08-17.

claude skill add Fission-AI/OpenSpec

For BAs working with engineering teams, this is also a translation layer. You produce a spec in a format the engineers already accept, and the arguments about whether a requirement was clear stop being arguments about memory.

PRD Taskmaster by anombyte93 generates product requirement documents and breaks them into tasks. 588 stars, community tier, unreviewed, MIT, updated 2026-08-14. It's newer and less proven than OpenSpec, and it's opinionated about PRD structure in a way you'll either like immediately or fight. Try it on a document you've already written and compare.

Data Analytics Skills by nimrodfisher bundles a set of analysis routines. 366 stars, community, unreviewed, and the license field reads Unknown, which is a genuine blocker in a corporate environment. If your employer has any software policy at all, an unknown license means legal says no. Worth watching, hard to recommend for work.

Trust tiers, briefly

Four of the ten skills here are community tier and unreviewed. That doesn't make them bad. It means nobody has checked them but you, and you should. The practical rule I use: knowledge-only skills that add context are low risk at any tier; anything that holds credentials, executes code, or touches files needs the source read first regardless of star count. We covered the reasoning in more depth in why skill author trust tiers matter.

Two of them have not been updated in ten months or more. Staleness in a skill that shells out to Python or connects to a database is a real hazard, because the dependency it pins may have moved on in ways that break quietly.

A setup that works

If you're starting from zero, install XLSX, DOCX, and PPTX first. They're official, audited, and they cover the output side of the job, which is where the hours go. Add MySQL MCP Server next if you have a database you're allowed to query, pointed at a read-only role.

Add OpenSpec only if you're producing specifications that engineers consume. If your deliverable is a deck for executives, it's the wrong tool and it will feel like overhead.

Leave the community-tier skills for last, install one at a time, and read the source for each. Your context window is finite, and every skill you add competes for room with the problem you're trying to solve.

For the tool side of the same job rather than the skill side, our full list for data analysts covers the platforms these skills sit alongside.

FAQ

Do I need Claude Code, or do these work in other agents?

Mixed. The Anthropic document skills are Claude Code first. The MCP servers here list Claude Code, Cursor, and in several cases VS Code Copilot as compatible. Check the compatible-agents field on each skill page before you commit to a stack.

Can these connect to Snowflake, BigQuery, or SQL Server?

Not the ones above. MySQL MCP Server is MySQL only. There are separate skills for Snowflake, ClickHouse, and SQL Server in the directory, with their own trust tiers and their own staleness problems. Search by your warehouse rather than assuming a general database skill covers it.

Is it safe to give an agent access to production data?

Give it a read-only role and a non-production replica, and the question mostly answers itself. The risk that matters isn't the model doing something malicious, it's a well-intentioned query locking a table during business hours. Treat the agent like a junior analyst with database access, because that's the correct threat model.

How many skills should I install at once?

Fewer than you want to. Each one consumes context, and an agent with fifteen skills loaded reasons worse about your actual question than one with four. Start with the three document skills, add the fourth only when you've hit a wall the first three didn't solve.

What about Excel MCP Server versus the XLSX skill?

They overlap and they're not redundant. XLSX is better for creating and analyzing files as part of a task. Excel MCP Server is better for repeated manipulation of a file that lives somewhere and stays there. Running both is fine if you actually use both, wasteful if you don't.