Best AI Agent Skills for Lawyers in 2026: 9 Skills Worth Installing

The legal-tech AI tools that flooded inboxes in 2024 (Harvey, Spellbook, Clio Duo) all share one shape: a hosted SaaS where you upload documents and get answers back. Claude Code skills invert that. The model runs against documents on your machine, with extensions you control, and nothing leaves your environment unless you tell it to. For lawyers handling privileged material, that distinction is the entire point.

This guide covers 9 Claude Code skills built for legal work in 2026. Each entry has the install command, what the skill is good at, where it falls down, and the security considerations a partner or compliance lead should sign off on before it touches a client matter.

For a tools-side view, see our full list of AI tools for lawyers.

What Counts as a "Legal Skill"

The useful skills cluster into four jobs:

1. Document handling — read, draft, compare, and redline Word/PDF 2. Citation and case research — find the case, pull the cite, check the holding 3. Privilege-aware research — keep matter-privileged content off third-party APIs 4. Compliance and audit — log what the model did and why, for ethics review

Every skill below sits in one of those buckets. Skip anything that promises "AI lawyering", that's a marketing layer, not a technical capability.

Quick Comparison

Document Handling

1. DOCX (Word Documents)

DOCX is the skill I'd install first for any litigator or transactional lawyer. It lets Claude read .docx files, parse the structure (headings, sections, comments, tracked changes), and write modified versions back. The track-changes preservation is the part legal practitioners actually need.

/install docx-word

Real use cases:

• Reading 80 pages of opposing counsel's draft and producing a clause-by-clause summary in 3 minutes
• Generating a redline against your firm's playbook and inserting tracked-changes comments
• Comparing two versions of a contract and producing a deviation report

Where it excels: Standard contract review work. The 200-page private placement memo that used to take a junior associate a full day gets a first pass in 20 minutes.

Where it falls short: Footnotes, complex tables of authorities, and embedded objects (Visio, Excel) sometimes don't round-trip cleanly. Always diff the output before sending. Also: the comments feature works, but Claude's tone in comment text reads slightly mechanical, humanize before sending to the client.

Security note: Operates on local files only. No outbound traffic unless paired with a research skill.

2. PDF Tools

PDF Tools handles the other half of legal practice, scanned filings, court orders, exhibits, and discovery PDFs. Includes OCR for scanned documents, annotation extraction, and page-level navigation.

/install pdf-tools

Use cases that pay back fast:

• OCR-ing a 400-page deposition transcript and answering questions across it
• Extracting all signature blocks from a folder of executed contracts
• Pulling exhibit references from a filed motion and matching them to the exhibit binder

Where it falls short: OCR accuracy on poor-quality scans (microfilmed cases, faxed exhibits) is good enough for keyword search but not for verbatim quoting. If you need to quote, verify against the source page image.

3. Doc Co-authoring

Doc Co-authoring is the skill for projects spanning multiple related documents, a litigation brief plus its exhibits and table of authorities, or a transactional close set with the credit agreement, security documents, and legal opinions.

/install doc-coauthoring

It's not the install I'd start with, DOCX and PDF cover most use cases, but for partners running large matters, the multi-doc context window saves real hours during weekly status updates and document reviews.

Research and Citation

4. Deep Research

Deep Research is the foundational research skill. It runs multi-step web research, returns sourced citations, and produces a structured memo with footnotes. For lawyers, the value is the citation discipline. Claude won't just synthesize, it'll point at the source for every claim.

/install deep-research

Where it earns its keep: drafting the background section of a research memo. Feed it the question, and you get a 4-page draft with 30+ citations in 12 minutes. You'll edit, but you're starting from a sourced position rather than a blank page.

Important caveat: Deep Research pulls from the open web, public legal databases, court websites, government sources. It is not a Westlaw or Lexis substitute. For paid-database research, you still need Westlaw, Lexis, or Bloomberg Law subscription access.

5. Agentset

Agentset chains multiple research sources together, a case law search, a secondary-source lookup, and a regulatory database query, into a single coordinated research task. For complex memos that pull from 5-10 source types, it's the difference between an hour and an afternoon.

The install is more involved (you'll configure source endpoints), and it's more powerful in the hands of a paralegal who's done research in the matter area before. New users will get over-broad results.

6. NotebookLM MCP

NotebookLM MCP brings Google's NotebookLM-style source-grounded Q&A into Claude. Upload a set of documents (deposition transcripts, expert reports, exhibits), and ask questions that get answered only from those sources, with passage citations.

/install notebooklm-mcp

Real use case: Discovery review for a mid-size matter. 4,000 pages across 60 docs. NotebookLM-style grounding gets you to the relevant 80 pages of the 4,000 in a couple of focused queries.

Privilege flag: NotebookLM's hosted version sends documents to Google. Confirm whether the MCP version is local-only or routes through the same hosted backend before using on privileged material. As of early 2026, several MCP versions exist with different routing, read the README.

Cross-Border Practice

7. Yargi MCP

Yargi MCP provides programmatic access to Turkish case law and statutes via the Yargi public legal database. Niche but valuable for firms with Turkish cross-border practice (energy, infrastructure, dispute resolution into Istanbul arbitration).

Setup is involved, you'll configure Yargi credentials and tune queries, and the value scales with how often Turkish law shows up in your work.

8. Korean Law MCP

Korean Law MCP is the equivalent for Korean statute and case lookups. Same pattern, niche, valuable for the firms with Korean practice, ignorable otherwise.

Worth flagging: there is no equivalent open MCP for U.S. case law that I'd recommend in production. The U.S. paid databases (Westlaw, Lexis, Bloomberg Law, vLex) don't expose stable MCP endpoints yet, and the open alternatives (CourtListener, Caselaw Access Project) cover only a subset and don't include up-to-date treatment. For U.S. work, MCP gives you a research assistant, not a citator.

Compliance and Governance

9. Claude Skills: Governance, Risk & Compliance

Claude Skills GRC is the install for in-house counsel, compliance teams, and law firms running their own SOC 2 or ISO programs. It packages compliance frameworks, control mappings, and audit-prep workflows into skills Claude can run against your evidence repository.

/install claude-skills-governance-risk-and-compliance

Use cases: drafting a SOC 2 control narrative from your evidence library, mapping NIST CSF to your existing policies, or generating a gap analysis for a new framework adoption.

Not for litigation or transactional work. But for the GRC-side of in-house legal, this is the most practical compliance skill currently published.

Skills I'd Skip in 2026

• Generic "summarize this PDF" skills, DOCX and PDF Tools cover this, with better fidelity
• "AI lawyer" or "draft any contract" skills, there's no shortcut around clause-level review
• Skills that scrape paid legal databases, they are an ethics complaint waiting to happen, and most violate the database's terms of service

The Install Order

If you're starting from zero:

1. DOCX (Word Documents), the highest-value install for any practicing lawyer 2. PDF Tools, to handle the half of your inbox that isn't Word 3. Deep Research, for memo drafting and background research 4. Doc Co-authoring, once you're regularly handling multi-doc projects 5. NotebookLM MCP, when discovery review enters the picture

The two cross-border skills and the GRC skill are situational. Install when the practice area calls for them, not before.

Security Considerations Before Installing Anything

Before any skill touches a client matter:

• Read the source. Skills are open code. Read what they do before installing, especially the network calls and the file access.
• Check the publisher. Anthropic-published skills (DOCX, PDF, Deep Research) are the safest baseline. Community skills require source review.
• Run on a test corpus first. Don't point it at the active matter folder on day one.
• Confirm where data goes. A skill that calls a third-party API to do its work has implications for privilege. The MCP version of NotebookLM is the example to scrutinize.
• Document for the file. If your firm has an AI use policy, log which skills you ran against which matter, and keep the prompts. Discovery requests are starting to ask for this.

For a deeper how-to on auditing, see our guide on auditing a Claude skill before installing.

FAQ

Q: Are Claude skills bar-rules-compliant?

The ABA Model Rules don't address AI skills specifically yet, but Comments to Rule 1.6 (confidentiality) and Rule 5.3 (supervision of nonlawyer assistants) cover the territory. Run skills with the same supervision discipline you apply to a junior associate or paralegal, and document it.

Q: Can I bill skill-assisted work to a client?

This is an ethics question for your jurisdiction. The mainstream view in 2026 is that skill-assisted work is billed at the human-effort rate, you bill your review and judgment time, not the skill's runtime. Confirm with your bar.

Q: Are these as good as Harvey or Spellbook?

Different shape. Harvey and Spellbook are end-to-end products with workflows baked in. Skills are building blocks Claude composes on demand. For a firm that wants a turnkey product, the SaaS tools are easier. For a firm that wants control over the data flow and workflow, skills are the right primitive.

Q: Do I need MCP and skills both?

You don't need to know the technical difference to use them. But for context: skills are reusable instruction packages, and MCP servers are connectors to external tools and data. Many of the entries above are technically MCP-backed skills. See MCP servers vs agent skills for the breakdown.

Q: What's the cheapest way to start?

Claude Pro plus the four core skills (DOCX, PDF, Deep Research, Doc Co-authoring) gets you a working setup for $20/month and 30 minutes of installation. Compare that to per-seat legal SaaS at $100-500/month and the math is straightforward.