Skip to main content
Guide9 min read·Updated June 11, 2026
🧩

Best AI Agent Skills for Virtual Assistants in 2026

B

A. Frans

Published June 11, 2026

AI Agent SkillsVirtual AssistantClaude CodeProductivityAutomation

A virtual assistant's job is mostly glue work: pull this into that, format this report, chase that document, summarize this inbox. It's exactly the kind of work agent skills were built to absorb. Skills extend Claude with specific capabilities, connect to a tool, run a workflow, remember context across sessions, so the assistant does the repetitive part and you do the part that needs a person.

Below are the skills worth installing if you run a VA practice or work as one. I've grouped them by the job they replace, with install commands and the security note that matters for each, because a skill that touches a client's Notion or inbox is a skill you need to trust.

SkillReplaces the chore ofWhere it runsRisk to watch
Notion MCPManual copy-paste into client docsClaude CodeWrite access to the whole workspace
FirecrawlTab-by-tab web researchClaude CodeScrapes any URL you point it at
Deep ResearchCompiling a research brief by handClaudePulls from open web, verify sources
Claude MemRe-explaining context every sessionLocalStores notes on your machine
Slack MCPDigging through channels for updatesClaude CodeReads message history
XLSX skillHand-building spreadsheetsLocalRuns on files you give it

Why skills beat a pile of subscriptions

A VA's stack tends to sprawl: a scheduling app, a research tool, a writing assistant, a transcription service, each its own login and bill. Agent skills collapse a lot of that into one place. Instead of five tabs, you tell Claude what you need and the right skill handles it. The work happens where you're already working, and the context carries between tasks, the research you ran feeds the doc you write feeds the spreadsheet you build, without you ferrying data between apps.

The trade-off is setup. Skills run mostly through Claude Code, which means a one-time install and a little comfort with a terminal. If that's a wall, start with the ones that don't need it (Deep Research runs in Claude directly) and add the rest as you get comfortable.

Notion MCP, the workspace connector

If your clients live in Notion, this is the first install. The Notion MCP lets Claude read and write pages, databases, and tasks directly. The VA use case is obvious the moment you have it: "summarize this week's client calls into the project doc," "create task cards from these meeting notes," "update the content calendar with these three posts." It does the structuring you'd otherwise do by hand.

Install it in Claude Code:

claude mcp add notion -- npx -y @notionhq/notion-mcp-server

You'll authenticate with a Notion integration token and grant it access to specific pages. That last part is the security note: give the integration access only to the pages it needs, not the entire workspace. A VA managing several clients should never connect one integration that can see every client's space. Scope it per client, and treat the token like a password.

Firecrawl, research without the 30 tabs

Half of VA research is opening tabs, skimming, and copying the useful bit. Firecrawl turns a URL (or a whole site) into clean text Claude can read and summarize. Point it at a competitor's pricing page, a list of event venues, a directory of prospects, and get structured output instead of a manual copy-paste marathon.

claude mcp add firecrawl -- npx -y firecrawl-mcp

It needs a Firecrawl API key (free tier available). The thing to watch: it'll scrape whatever you point it at, so be sensible about terms of service and rate limits on sites you don't own. For client research across public pages, it's a real hour-saver. Check the GitHub repo (github.com/mendableai/firecrawl) for current setup and the API key step.

When a client asks "find me everything on X," Deep Research is the skill that runs the multi-source sweep and hands back a cited brief instead of a list of links. It fans out across the web, reads sources, cross-checks claims, and writes a summary you can hand off after a review.

This one runs in Claude directly, no terminal needed, which makes it the best starting point for VAs who aren't technical. The discipline it requires from you: verify the sources before you forward anything to a client. The skill cites where it got each claim, open the important ones and confirm them. AI research is a strong first draft, never a final word, especially on anything a client will act on.

Claude Mem, stop re-explaining the client

Every new Claude session starts blank. For a VA juggling five clients, that means re-explaining each client's preferences, tone, and quirks constantly. Claude Mem (and similar memory skills) fixes that by storing context locally and recalling it when relevant. Set up a memory of "Client A prefers short bullet updates, no emoji, sends invoices on the 1st" once, and it's there next week.

# Install via the claude-mem plugin, see the GitHub README for current steps
npx claude-mem install

The security upside here is real: it stores memory locally on your machine, not on someone's server, so client preferences stay with you. Keep genuinely sensitive details (passwords, financials) out of any memory store regardless, memory is for working context, not secrets. Confirm the exact install command on the project's GitHub before running it, since memory tooling moves fast.

Slack MCP, the standup nobody has to attend

For VAs embedded in a client's team, Slack is where the work updates live and where hours disappear scrolling for context. The Slack MCP lets Claude read channel history and summarize it: "what did the marketing channel decide about the launch date," "draft a reply to the client's question in #general." It reads the firehose so you don't.

claude mcp add slack -- npx -y @modelcontextprotocol/server-slack

This one reads message history, which is the privacy line to be careful with. Use it only on workspaces where you're an authorized member and the client knows an assistant tool has read access. Don't connect a client's Slack without saying so, that's a trust violation waiting to happen, regardless of how useful the summaries are.

The XLSX skill, spreadsheets without the grind

A lot of VA deliverables are spreadsheets: expense trackers, content calendars, contact lists, simple reports. The XLSX skill builds and edits real .xlsx files, formulas, formatting, charts, from a plain description. "Turn this CSV of leads into a sorted, formatted tracker with a status column" becomes one instruction instead of twenty minutes of clicking.

It runs locally on files you hand it, which is the safe model: nothing leaves your machine. It's bundled with Claude's document skills, so check whether it's already available in your setup before installing anything. For VAs who live in spreadsheets, it's one of the biggest time-savers here.

How to roll these out without breaking things

Don't install all six on day one. Start with the skill that targets your biggest time sink, usually Notion MCP if clients live there, or Deep Research if you do a lot of research. Get one workflow solid, learn its quirks, then add the next.

Two rules hold across all of them. First, scope access tightly: per-client tokens, least privilege, never one connection that sees everything. Second, keep a human review step on anything client-facing. These skills draft, summarize, and structure fast. They also make confident mistakes, and a wrong figure in a client report is your name on it, not the tool's. The VAs who win with skills are the ones who use them to do twice the work, not the ones who stop checking.

For the non-skill side of the stack, scheduling, transcription, the apps that don't run through Claude Code, see our full list of AI tools for virtual assistants.

FAQ

Do I need to know how to code to use these skills?

For most, no, but you need to be comfortable pasting a command into a terminal once to install the MCP-based ones. After setup, you interact in plain English. Deep Research and the document skills run inside Claude with no terminal at all, so start there if a command line feels intimidating.

Are agent skills safe to use with client data?

They can be, with discipline. Use local-first skills (memory, spreadsheets) for sensitive work since nothing leaves your machine, scope every connected tool to the minimum access it needs, and tell clients which tools have read access to their workspaces. The risk isn't the skill, it's connecting it too broadly or without disclosure.

How much do these skills cost?

The skills themselves are mostly free and open source. Costs come from the services behind them, a Firecrawl API key, a Notion plan, your Claude subscription. Several (Deep Research, Claude Mem, the XLSX skill) add no extra cost on top of Claude. Check each tool's GitHub or pricing page for the current details.

Can one skill manage multiple clients?

Technically yes, but scope them per client for safety. Run separate Notion integration tokens and separate memory contexts per client so one tool never sees across client boundaries. It's a little more setup and a lot less risk if a client ever audits what your tools can access.

Where do I find the official install commands?

Always confirm the command on the skill's GitHub repository before running it, these tools update often and a command from a six-month-old blog post may be stale. Search the skill name plus "GitHub," check the README, and prefer the official repo over third-party instructions.

Share this article

📬

Get More AI Tool Guides

New comparisons and guides every week. Join thousands of professionals staying ahead of the AI curve.