Best AI Agent Skills for Data Analysis in 2026

A spreadsheet with 80,000 rows used to mean an afternoon of pivot tables and a headache. Now you point Claude Code at the file, it loads the right skill, writes the analysis, runs it, and hands you the chart. The difference between a skill that does this well and one that wastes your tokens flailing is real, so here are the agent skills I'd actually install for data work, what each one is for, and the security check you do before any of them touch your data.

What an "agent skill" is, quickly

A skill is a folder with a SKILL.md file that teaches Claude a specific capability: how to parse a spreadsheet, how to query a database, how to build a chart. Claude loads it only when the task calls for it, so you can install a dozen and only pay the context cost of the one in use. That on-demand loading is why a stack of data skills doesn't bloat every conversation.

For data analysis specifically, the useful skills fall into three jobs: getting data in and out of files, computing over it, and visualizing the result.

The skills worth installing

xlsx-spreadsheet: the one you'll use most

If you do data work, this is the first skill to install. Business data lives in spreadsheets, and this skill lets Claude open an .xlsx or .csv, understand the columns, run the calculation you asked for, and write results back into a clean file with formulas intact. It's the difference between Claude describing what you should do in Excel and Claude just doing it.

The honest limit: very large files still stress any agent, and you'll get better results pointing it at the specific sheets and ranges you care about rather than dumping a 50-megabyte workbook and saying "analyze this." Scope the ask and it's reliable.

deep-research: context your spreadsheet doesn't have

Your data tells you what happened inside your business. It can't tell you whether a 4% conversion rate is good for your industry. deep-research fills that gap by running structured web searches, fetching sources, and giving you cited benchmarks to compare against. I pair it with a spreadsheet analysis constantly: compute the internal number, then ask deep-research what the industry baseline is, and the comparison is where the actual insight lives. We reviewed it on its own in our [deep-research skill review](/blog/deep-research-skill-review-worth-installing-2026) if you want the deeper take.

d3js-visualization: making the result legible

Numbers in a table don't persuade anyone. d3js-visualization builds proper charts and interactive visualizations from your analysis, which matters the moment you have to present findings to someone who won't read a spreadsheet. It's more capable than the basic chart any tool can spit out, and it's worth the install if part of your job is communicating results, not just computing them.

pdf-tools: freeing trapped data

So much useful data sits inside PDF reports, financial statements, and exported dashboards that nobody bothered to also publish as a CSV. pdf-tools extracts the tables and text so you can actually compute over them. The extraction isn't magic on badly scanned documents, but for digital PDFs with real tables, it gets the data out cleanly enough to work with.

mcp-builder: when the data is in a database

Files are one thing. If your real data lives in Postgres or another database, you want Claude talking to it directly, and that's an MCP connection rather than a skill in the strict sense. mcp-builder helps you set up that bridge so Claude can query live data instead of you exporting to CSV every time. For analysts whose source of truth is a warehouse, this is the unlock.

How to install

Skills install into Claude Code's skills directory. The pattern is the same for any of them: clone the skill repo into your skills folder, and Claude picks it up automatically.

``bash # Claude Code looks for skills in ~/.claude/skills/ cd ~/.claude/skills/

# Clone the skill you want git clone https://github.com/anthropics/skills.git

# Or, for a single skill, copy just its folder into ~/.claude/skills/ # Each skill is one folder containing a SKILL.md file ``

Restart Claude Code or start a new session, and the skill is available. Claude loads it on its own when a task matches the skill's description, so you don't invoke it manually for routine work.

The security check you do first

This is the part people skip and shouldn't. A skill is code and instructions that run with your permissions, against your data. Before you install one, especially from a source that isn't Anthropic's official repo, do three things.

Read the SKILL.md. It's plain text. You can see exactly what the skill tells Claude to do, including any commands it runs and any external services it calls. If it wants to send your data somewhere, you'll see it.

Check the repo. Look at the GitHub source, the star count, the open issues, and when it was last updated. A skill with one commit from an account with no history is a different risk than one from a known maintainer with a real track record.

Watch for data exfiltration. The specific thing to look for: does the skill upload your files, send content to a third-party API, or phone home? For data analysis that's a hard line, because your spreadsheets often contain the most sensitive information you own. If a skill needs network access, it should be obvious why.

The official Anthropic skills repo is the safe default. Community skills can be excellent, but you audit them first. We wrote a fuller guide on [why skill security matters and how to audit them](/blog/why-ai-agent-skill-security-matters-how-to-audit-2026), and for data work the stakes are higher than usual because the input is your private numbers.

A realistic data workflow

Put together, a normal analysis session looks like this. You drop a messy CSV export and ask Claude to clean it; xlsx-spreadsheet handles the parse and the dedupe. You ask for the actual analysis, segment revenue by month, find the outliers, whatever the question is, and it computes and writes the result to a new file. You ask how your numbers compare to your industry; deep-research pulls cited benchmarks. Then you ask for a chart you can drop into a deck, and d3js-visualization builds it. Four skills, one conversation, and the part that used to be an afternoon is now the time it takes to read the output.

That's the case for installing a small, deliberate set of data skills rather than one do-everything tool. Each does its job well, Claude only loads what the task needs, and you keep control over what touches your data.

Skills or a dedicated BI tool?

If you already pay for Tableau or Power BI, skills still earn a spot, because the two are good at different things. A BI tool is built for dashboards you refresh and share with a team, the same charts updating on a schedule. Agent skills are built for the one-off question, the messy file nobody has a dashboard for, the "just tell me what's weird in this export" task that would take 20 minutes of manual pivoting.

The two coexist well. Use the BI tool for the metrics you track forever. Use Claude with these skills for the ad-hoc digging, the data cleaning, and the questions you'll ask once and never again. In practice the skills handle the long tail of analysis that never justified building a proper dashboard, which for most analysts is a surprising share of the actual workload.

One cost note: each skill runs real computation, so a heavy analysis over a large file uses more tokens than a quick chat. Scope your asks to the data you care about and you'll keep usage reasonable.

FAQ

Do I need to be a programmer to use these skills? No. The whole point is that Claude writes and runs the code. You describe what you want in plain language. Installing a skill takes one git clone command, and you can copy that exactly from the examples above.

Which skill should I install first for data analysis? xlsx-spreadsheet. Most business data is in Excel or CSV, and this skill covers the largest share of everyday analysis tasks before you add anything else.

Are these skills free? The skills themselves are open source and free. You pay for the Claude usage that runs them, the same as any Claude Code session. There's no separate skill fee.

Can a skill see my data? Yes, that's how it works, which is why you audit the SKILL.md before installing. A skill runs with your permissions against the files you point it at. Stick to the official repo or skills you've read, and never run one that sends your data to an unexplained external service.

Do these work outside Claude Code? Agent skills are a Claude Code feature. The underlying capabilities exist elsewhere, but the install-a-folder, load-on-demand model described here is specific to Claude Code's skills system.