Claude for Legal: Practice-Area Plugins, Contract Review, and Lawyer Workflows

Claude for Legal is Anthropic's open-source collection of practice-area plugins, skills, scheduled agents, and MCP connectors for legal work. You install the plugins in Claude Cowork or Claude Code, run a short setup interview that learns your playbook, then use commands like /commercial-legal:review for vendor agreements and NDAs. Anthropic's own repo is blunt about the limit: every output is a draft for attorney review, not legal advice. This guide covers what ships, a setup workflow, and the confidentiality questions to settle first.

TL;DR: 12 practice-area plugins plus 20+ legal MCP connectors (Ironclad, Docusign, iManage, Everlaw, Thomson Reuters CoCounsel and more). Install in Cowork (Customize → Browse plugins) or Claude Code, run the cold-start interview, connect a research tool before trusting citations, and keep a lawyer accountable for anything that leaves the building. Checked 2026-10-12 against Anthropic's announcement, the anthropics/claude-for-legal README, the Usage Policy, and the Privacy Center.

What Anthropic actually ships

Per Anthropic's post Claude for the legal industry (checked 2026-10-12), the legal release has three building blocks:

• MCP connectors that bring matter documents, communications, and records into Claude. Anthropic lists 20+ new ones across contract lifecycle (Definely, Docusign, Ironclad), deal rooms (Box, Datasite), document management (iManage, NetDocuments), e-discovery (Consilio, Everlaw, Relativity), research (Thomson Reuters CoCounsel Legal, Legal Data Hunter, Midpage, Trellis), legal AI assistants (Harvey, Solve Intelligence), and public-service tools (BoardWise, Courtroom5, Descrybe, Free Law Project).
• 12 practice-area plugins: Commercial, Corporate, Employment, Privacy, Product, Regulatory, AI Governance, IP, and Litigation Legal, plus Law Student, Legal Clinic, and Legal Builder Hub.
• Open protocols, so firms and in-house teams can add their own precedents, playbooks, and skills.

LawNext reported the release on May 12, 2026, building on a first legal plugin for Cowork in February. Anthropic says the connectors and plugins are open source and available in Claude Cowork, and that Enterprise admins enable them in workspace settings. A plan-by-plan availability table for the legal plugins is not published as of 2026-10-12; check Claude pricing for what your plan includes.

The code lives in the public anthropics/claude-for-legal repo (Apache 2.0). Each plugin has a template CLAUDE.md practice profile, skills you call as /<plugin>:<skill>, and in some cases scheduled agents such as a renewal watcher or docket watcher.

When to use Claude for legal work / when not

Use it when:

• You review high volumes of routine agreements (NDAs, vendor MSAs, DPAs) against a playbook you already have.
• You need first-pass structure: chronologies, diligence grids, privilege-log flags, closing checklists, renewal registers.
• A lawyer will review, verify, and take responsibility for the output.
• You can connect a research tool so citations are checked against a database, not model memory.

Don't use it when:

• Nobody qualified will review the output before a client relies on it.
• You want a legal conclusion or opinion. The repo calls FTO and infringement outputs "triage, not an opinion."
• Your firm hasn't decided whether client material can go into Claude, under which plan and retention settings.
• You need citations you can file and no research connector is set up.

Workflow: from install to first contract review

1. Pick the plugin that matches your seat. In-house commercial work starts with commercial-legal. M&A teams start with corporate-legal. Litigators start with litigation-legal.
2. Install it. In Cowork: open the Cowork tab → Customize → Browse plugins. In Claude Code, the README shows /plugin marketplace add with the repo path, then /plugin install commercial-legal@claude-for-legal.
3. Run the cold-start interview. /commercial-legal:cold-start-interview. The README says it takes 10–20 minutes per plugin and asks for seed documents (a signed MSA, a playbook, a prior review memo). It writes your practice profile, and every other skill reads from it. Skipping it is, per the README, "the single most common reason a skill produces generic output."
4. Connect a research tool. The README says to do this first. Without it, citations from model knowledge are flagged [verify] and the deliverable notes that sources weren't verified.
5. Run one real job. For example, /commercial-legal:review on an inbound NDA gives a GREEN/YELLOW/RED triage, so only the hard ones reach a lawyer.
6. Review in Word. The README says every contract-touching skill is written for the Claude for Word sidebar, with tracked changes as the output. You accept or reject each change like a human markup. See our Claude for Word guide.
7. Tune the profile. Edit CLAUDE.md when an escalation threshold or fallback clause is wrong. Re-run the interview when your practice changes a lot.

Jobs worth starting with

Job	Plugin · command	What you get
NDA / vendor MSA triage	commercial-legal · /commercial-legal:review	Playbook-based review and redline memo
Renewal deadlines	commercial-legal · /commercial-legal:renewal-tracker	Contracts with cancel-by dates within 90 days
DPA review	privacy-legal · /privacy-legal:dpa-review	Review as controller or processor
Diligence grid	corporate-legal · /corporate-legal:tabular-review	One row per document, every cell cited
Chronology	litigation-legal · /litigation-legal:chronology	Chronology from declared sources and uploads
Privilege log	litigation-legal · /litigation-legal:privilege-log-review	Obvious calls plus flags for attorney review

All commands above are as listed in the repo README on 2026-10-12.

The Claude moves worth stealing, by email

Workflows tested on real jobs, sent only when they are worth it. Free.

Subscribe

Claude for lawyers: what changes in practice

If you're a practising lawyer, three things matter more than the feature list.

Accountability stays with you. The README says the plugins "do not represent Anthropic's legal positions" and that "the attorney using the plugin — not the plugin, and not Anthropic — is responsible for the legal positions taken in their work product." Anthropic's Usage Policy treats advising an individual on their legal rights, obligations, or liability, or preparing legal documents for them to sign or file, as a high-risk use. That requires a qualified person (licensed where the law requires it) to review the output before it is delivered, and the person receiving the advice must be told AI was used. Internal drafting, research, and summarization that isn't the final advice is excluded from those requirements.

Citations need a source. Connect CourtListener, Trellis, Descrybe, CoCounsel, or another research connector, and treat anything flagged [verify] as unchecked.

Confidentiality is your call to make, not Claude's. Anthropic's materials describe connectors that respect existing permissions (Ironclad results scoped to user permissions, iManage "permission-bound, auditable access"). They do not say how privilege applies when you use them; Anthropic does not publish a privilege position as of 2026-10-12. On retention, Anthropic's Privacy Center says that on consumer plans (Free, Pro, Max), deleted chats leave back-end storage within 30 days. If you allow model training, data may be kept de-identified for up to 5 years. Commercial products such as Claude for Work and the API are covered by separate terms. Settle your plan, training setting, and firm policy before any client file goes in. Use your bar's guidance and your engagement terms; this guide is not legal advice.

Small firms weighing cost should read Claude for small business and Is Claude free? alongside this.

Prompts

Three copy-paste prompts to start with:

1. Playbook NDA triage (no plugin)

You are assisting a lawyer. Review the attached NDA against the playbook below.
For each clause: GREEN (matches), YELLOW (acceptable fallback), RED (escalate).
Quote the clause, cite the playbook rule, and propose fallback language only
from the playbook. Do not give a legal conclusion. End with open questions.
Playbook: [paste]

2. Chronology from documents

Build a chronology from the attached documents only. One row per event: date,
event, source document and page, confidence. Flag undated or conflicting
entries. Do not infer facts not in the documents. Mark privileged material
for attorney review instead of summarizing it.

3. Research plan, not answers

I'm researching [issue] in [jurisdiction]. Give me a research roadmap: statutes
to check, case law areas, and search terms. Treat every authority as a lead to
verify, not a citation. Flag where the law is unsettled.

Agent block

Role: legal drafting assistant for [firm/team]. Every output is a draft for
attorney review, not legal advice. Use only provided documents and connected
research sources; flag any citation from memory as [verify]. Never send, file,
or share anything. Stop and ask before using client data outside this matter.
State jurisdiction assumptions. Mark privilege calls for attorney decision.

Stack

• Chat / Projects: Anthropic says Projects give matter teams a persistent workspace where precedents and prior drafts carry across conversations.
• Cowork: batch work across many documents and scheduled tasks (weekly regulatory sweeps, intake triage).
• Claude Code: same plugins, installed via the plugin marketplace.
• Microsoft 365: Word for redlines, Outlook for matter intake and cover notes, Excel for diligence and claim charts.
• Skills: the plugins' own skills, plus community skills through Legal Builder Hub, which runs security, license, and freshness checks before install.
• Managed Agents: a subset ships as cookbooks for programmatic deployment (diligence grid, docket watcher, launch radar, reg monitor, renewal watcher).

Pitfalls

• Skipping the cold-start interview. You get generic output.
• Trusting uncited authority. No connector means unverified cites.
• Installing random community skills. The README warns they run with access to your matter files; use the hub's trust gate.
• Unclear data posture. Pick the plan and training settings before uploading client documents.
• Letting a draft leave unchecked. The plugins include gates before anything is filed or sent. Don't route around them.

FAQ

Is Claude for Legal a separate product? No. It's a set of open-source plugins, skills, and connectors that run in Claude Cowork and Claude Code, and can be deployed through Managed Agents.

Does it give legal advice? No. Anthropic's repo says outputs are drafts for attorney review, not legal advice or legal conclusions.

Which plans include it? Anthropic says it's available in Cowork and that Enterprise admins enable it in workspace settings. A full per-plan list is not published as of 2026-10-12.

Can it check citations? Yes, with a research connector. Without one, citations from model knowledge are flagged [verify].

Is it free? The plugins are Apache 2.0. Many connectors (Ironclad, iManage, Everlaw, CoCounsel) need your own subscription, per the README.

Sources checked 2026-10-12: Anthropic, "Claude for the legal industry"; anthropics/claude-for-legal README; Anthropic Usage Policy; Anthropic Privacy Center, "How long do you store my data?"; LawNext, May 12, 2026.