Claude for Legal adds practice-area plugins, legal-tool connectors and a set-up interview to a normal Claude plan, so Claude works from your firm's playbook instead of generic instructions. A small firm should turn on one plugin for its commonest matter type, usually commercial-legal or employment-legal, trial it on five closed matters, and leave the rest off.
The word "add" matters. The model underneath is the same Claude you may already use. What changes is the packaging: workflows for specific legal tasks, a profile of how your firm works, and connections to document and research systems. That is useful, and it also widens what the tool can reach, which is why the order in which you turn things on is the real decision.
What Claude for Legal adds, in plain terms
Anthropic announced the legal pack on 12 May 2026, in a post titled "Claude for the legal industry" that introduced 12 practice-area plugins and more than 20 new connectors. Its legal page describes plugins for contract review, NDA triage and compliance workflows, configurable to an organisation's playbook; connectors to systems such as iManage, NetDocuments, Ironclad and Bloomberg Law, with DocuSign, Everlaw, Box and Thomson Reuters among the named partners; and Claude working inside Word and Outlook. The page also describes ready-made legal artifacts for tabular review, regulatory tracking, playbook review, matter tracking and legal intake. The plugins live in an open repository, anthropics/claude-for-legal, which firms can read before installing anything.
Compared with an ordinary Claude plan, you get four things. First, task workflows that know the shape of a legal job, such as a contract review that checks against a playbook and escalates exceptions. Second, a firm profile created by a short interview, so output reflects your positions. Third, connectors that let Claude read documents and research sources you already use. Fourth, guardrails the repository describes: source attribution on citations, conservative defaults on privilege and subjective judgements, surfaced jurisdiction assumptions, and gates before anything is filed, sent or relied on.
What you do not get is a substitute for supervision. The repository states that every output is a draft for lawyer review and is not legal advice, and that the lawyer remains responsible for the positions taken. Read that as the design brief: the tool prepares, the lawyer decides.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowPlans and price: what is and is not published
Anthropic's legal page does not show a per-seat price for Claude for Legal. It refers to Team and Enterprise plans and says that on those plans Anthropic does not train on your content by default. Anthropic's pricing page lists Team seats from two seats, with a Standard seat at $20 a seat a month billed annually or $25 monthly, and a Premium seat at $100 or $125, with Enterprise from $20 a seat a month annually with a 20-seat minimum and usage billed at API rates on top. Pro, for one person, is $17 a month billed annually or $20 monthly.
Two cautions. First, those are list prices on the page in November 2026, and Anthropic says prices and plans are subject to change. Second, the plugins carry no separate charge on Anthropic's pages: its help centre says plugins are available on all paid plans (Pro, Max, Ultra, Team and Enterprise), and the launch post says the legal plugins and connectors run in Claude Cowork, with Enterprise admins enabling them in workspace settings. What costs extra is the systems behind the connectors. The repository marks many of them, Everlaw and Solve Intelligence among them, as needing the customer's own subscription and API key, and the Thomson Reuters CoCounsel plugin needs a CoCounsel Legal subscription. Add those to your cost sum. For client work, Team rather than individual Pro seats is the sensible floor, because Team adds admin controls over connectors and no training on your content by default.
For a small firm, a reasonable starting budget is Team Standard seats for the fee earners who will trial it, not the whole firm. A sum for an illustrative five-person firm with three trial seats: 3 seats at $20 billed annually is $60 a month, or $75 on monthly billing. That compares with the time value of even a few saved hours, but only if the trial shows time saved after review. The wider pricing comparison is in per-seat or pay-as-you-go pricing for small firms.
The plugins, sorted for a small firm
The repository lists these practice areas. Here is how I would sort them for a firm of two to ten people, with the caveat that your matter mix decides.
| Plugin | What it covers (from the repository) | Small-firm verdict |
|---|---|---|
| commercial-legal | Playbook-aware review of vendor agreements, NDAs and SaaS terms; amendment tracing; renewal alerts | First choice for most firms with business clients |
| employment-legal | Hire and termination review, worker classification, leave tracking, investigations, policy drafting | Strong second if employment is a core area |
| corporate-legal | Diligence with tabular review, disclosure schedules, closing checklists, written consents, entity compliance | Useful for transactional firms; heavier to set up |
| litigation-legal | Chronologies, claim charts, deposition prep, privilege logs, brief drafting | Trial only where you have disciplined file handling |
| ip-legal | Trademark clearance, freedom-to-operate triage, takedowns, open-source compliance | Niche; turn on if IP is a regular work type |
| privacy-legal | Privacy triage, impact assessments, DPA review, subject-access responses | Good if you advise on data protection |
| product-legal, regulatory-legal, ai-governance-legal | Launch review, regulatory monitoring, AI use-case triage | Aimed at in-house teams; usually later for a small firm |
| legal-clinic, law-student | Clinic intake and student coaching | Not for a commercial practice |
| legal-builder-hub | Finds and vets community skills, with a trust gate | Leave off until you have a policy for third-party skills |
The repository also lists an external plugin from Thomson Reuters for its CoCounsel research product, which requires that subscription and is maintained by the vendor. Plugin names and contents may change after I wrote this, so open the repository and read the current README before you choose.
Why "one plugin, one matter type" is the right trial
Each plugin you install reads from your practice profile and, where connected, your files. Every additional plugin adds scope to review, instructions to test and connectors to govern. A firm that turns on six in week one cannot tell which one produced a bad result. A trial of one plugin on one matter type gives a clean signal.
The repository says to run the cold-start interview first: every skill reads from the practice profile it writes, and skipping it is the single most common reason for generic output. The interview takes 10 to 20 minutes per plugin, writes the profile to a plugin configuration folder on your machine, and asks you to point it at seed documents such as a signed agreement, a playbook or a past review memo. Those seed documents pass through Claude like any other file, so until your agreement has been reviewed, use anonymised or firm-owned precedents.
The running example: a six-person firm trialling commercial-legal
Consider an illustrative six-person firm, two partners, two associates, a paralegal and an office manager. Most of its work is business clients: supplier agreements, NDAs, software subscriptions and the occasional employment matter. It chooses commercial-legal, because a typical week brings six to ten NDAs and a couple of supplier contracts, and because the same review points recur.
- Week one: governance. A partner reads the repository README, Anthropic's legal page and the firm's plan terms. She records the decision on plan, connectors and who administers the workspace. Connectors to the firm's own document systems stay off.
- Week one: set-up. An associate installs the plugin and runs the cold-start interview with the firm's standard NDA playbook: acceptable term length, governing-law preferences, carve-outs, positions that need partner approval. Seed documents are the firm's own precedent NDAs with names removed.
- Weeks two and three: dry run. Five closed NDAs, anonymised, run through the plugin. The associate compares its issue list and proposed redlines with what the firm actually did at the time.
- Week four: decision. A short meeting reviews the numbers and chooses to proceed on live NDAs with review, adjust the playbook or stop.
What a first NDA review looks like, and what to correct
Here is an illustrative, trimmed issue list of the kind a playbook review produces for an inbound mutual NDA once the profile is loaded. The exact layout of your output will differ; the review habits are the point.
NDA REVIEW - draft for lawyer review | Counterparty paper | Playbook v1
1. Term (cl. 3): 5 years from disclosure. Playbook max: 3 years.
DEVIATION - propose 3 years.
2. Residuals (cl. 7): lets the recipient use information "retained
in unaided memory". Playbook: not acceptable. ESCALATE to partner.
3. Governing law (cl. 12): counterparty's choice. Playbook: ours
preferred. Jurisdiction assumption stated: [assumption].
4. Non-solicit (cl. 9): 12 months. Playbook: silent. FLAG.
5. Return or destroy (cl. 5): no carve-out for automatic backups.
Propose standard carve-out.
6. Remedies: injunctive relief note cites an authority [verify].An associate would make three corrections before anything reaches a partner. Item 6 carries a [verify] tag, so the authority came from model knowledge and leaves the memo unless someone finds it in a research source. Item 4 exposes a gap in the playbook, not a problem in the NDA: decide the firm's position on non-solicits once and add it to the profile, or the plugin will flag the same clause on every NDA. And every clause number gets checked against the document, because a wrong cross-reference in a note to a client undermines everything else in it. Recording which items were right, wrong or missing on each of the five matters gives you the trial numbers below.
The measures are simple: issues the plugin found that the firm also found, issues it missed, false flags, and minutes to a reviewed first draft. An illustrative result: on five NDAs, the plugin matched 22 of 26 issues the firm had found, missed 4, raised 6 false flags, and cut first-draft time from about 35 minutes to 15 including review. That would justify a cautious live trial. A result where it missed a one-sided indemnity would be a reason to rework the playbook first.
Connectors: add them in a second step
Connectors read, and sometimes write, client data in systems like document management, storage and e-signature. The repository says a connector gives Claude the ability to read from and, where scoped, write to your data, and it marks many as needing the customer's own subscription and API key. Treat them as a separate decision from the plugin.
A sensible order for a small firm: first a research connector, which the repository recommends connecting before anything else because citations are flagged as unverified without one, and which searches case-law sources rather than your files (CourtListener, from the Free Law Project, is one it lists); then one read-only connector to the place precedents live; and only later anything that writes or sends. For each, record which folders or matters are visible, who authorised it, and how to revoke it. On a Team workspace, an owner adds a connector for the whole organisation in the admin settings, each person then signs in to it individually, and owners can restrict what a connector may do, for example allowing reads while blocking writes. That gives you a natural control point.
Also be careful with the community-skills route. The repository warns that nobody certifies community skills, and that installing one means installing code that runs with access to your matter files, your practice profile and your research connectors; the builder hub adds a trust check before anything installs. Until you have a written rule for third-party skills, do not install any.
Privilege, confidentiality and the checks that stay human
Anthropic's legal page says that on Team and Enterprise plans it does not train on your content by default, that Enterprise admins set how long content is retained and that a lawyer remains in charge of the work. The commercial-legal plugin's notes add that privilege markers are applied conservatively and that consequential actions such as filing, sending or executing need explicit confirmation. Neither source answers the question a firm actually has to settle, which is whether using the service on privileged material fits your professional duties and your client terms. That answer sits in your agreement with Anthropic and your regulator's guidance, which you must read.
Keep these human checks in every workflow: read every cited authority in the source; confirm the jurisdiction the plugin assumed; check every figure, date and defined term; and keep a record of who reviewed. Where output is marked [verify], the citation came from model knowledge alone, so it should not travel into any document until you have found it in a research source.
For a worked contract workflow, see Claude legal review of contracts for a small firm. For comparing assistants on one legal task, which AI is best for legal work gives a side-by-side. Staffing and rollout questions are in the first-90-days plan for a small law firm, and the broader picture in the AI guide.
What to ask Anthropic in writing before the trial
Marketing pages answer most questions in outline and none in the detail your engagement terms need, so send a short list and keep the replies on file. Ask which plan tier each plugin and connector requires, and what each costs per seat. Ask how long prompts, outputs and uploaded files are retained on your plan, and whether an administrator can change that. Ask where data is processed, whether any human at the vendor can see conversations and under what conditions, and how a leaver's data is handled when an owner removes a seat.
Ask about privilege directly: does the agreement say anything about preserving privilege when a lawyer uses the service, and does Anthropic publish any guidance for law firms on privilege and confidentiality? Ask how changes to plugins and connectors are announced, since a behaviour change between versions is a risk to a tested workflow. If an answer is vague, do not fill the gap with optimism; record it as an open risk and keep the trial to anonymised material.
A short worked example shows why it matters. Imagine two partners disagree on whether a connector to the document store is safe. One reads the vendor page and says yes; the other points out that the page says nothing about retention for connected content. The written reply settles it, and the firm either enables the connector for one folder or leaves it off. Without the reply, the decision rests on tone, and tone is a poor basis for a client-confidentiality call.
When a normal Claude plan is enough
You do not always need the legal pack. If your use is occasional, such as rewriting your own paragraphs in plain English, structuring file notes, or building checklists, a normal Claude plan with a Project holding your house style may do. The plugin pays off when you repeat the same structured review often enough that a stored playbook saves real time, for example ten or more NDAs a month, a regular flow of vendor contracts, or recurring employment policies.
A rough break-even sum: if a plugin-assisted first review saves 15 minutes on each of 30 NDAs a month, that is 7.5 hours. At an internal cost of $60 an hour, that is $450 of time against a seat cost of $20 to $25. The sum looks one-sided, so the real question is whether review time stays honest: if reviewers rubber-stamp, the saving is an illusion and the risk is real. Measure the review minutes as well as the drafting minutes.
Four more matter types, quick verdicts
An illustrative estate-planning practice has little use for most plugins: its work is bespoke drafting and client meetings, so a plain Claude Project for plain-English letters is enough. A firm doing mostly residential property has no matching plugin; the checklist approach in a Project fits better. A small employment practice gets real value from the employment plugin's policy drafting and classification review, provided every output is checked against current local rules by a lawyer. A boutique IP firm may find the trademark clearance and takedown workflows useful, with research connectors, but must verify every authority.
A one-page decision record for the partners
- Plan and seats, with the date prices were checked on Anthropic's page.
- Plugin chosen, matter type and why.
- Connectors: research only in the dry run; the first document connector named, with scope and owner.
- Data rule: anonymised or firm-owned material only until the agreement has been reviewed.
- Review rule: named reviewer for every output; citations read at source.
- Trial: five closed matters, the four measures, the decision date.
- Stop rule: any missed material issue triggers a playbook review before further use.
That page is the real deliverable of the first month. It turns a launch announcement into a governed experiment, and if the plugin disappoints you will know why, with numbers.
Keeping the playbook current
A plugin that reads from a playbook is only as current as that playbook. Give it an owner and a review date, and update it whenever the firm changes a standard position or a client-specific exception becomes the norm. After the trial, keep the comparison file: the five matters, what the firm found, what the plugin found. Re-run it when a plugin updates, because new versions can change behaviour. The repository gives a one-line update command, so someone should own the update step as well, and should re-run the five-matter comparison after each update.
A final practical note. Tell clients how you use AI in your engagement terms, in a sentence or two: that AI tools may assist with drafting and review, are used on a business workspace, and are always supervised by a lawyer. Many clients will ask, and a clear answer builds trust. If a client asks you not to use AI on their matter, record that and respect it, and make sure the person opening the plugin knows which matters are excluded before they start.
Claude for Legal questions from small firms
Is Claude for Legal a separate product with its own price?
No separate price is published. Anthropic's legal page shows no price, and its help centre says plugins are available on all paid plans: Pro, Max, Ultra, Team and Enterprise. You pay the normal seat price, plus any subscription a connected legal system needs. Confirm the plan your firm needs with Anthropic before you commit.
Does Claude replace a legal research database?
No. The plugin pack tags citations that come from model knowledge alone with a verify marker, and research connectors, such as case-law sources, tag theirs by source. You still read the authority yourself before relying on it or citing it.
Can I use it on privileged client material?
Only after checking the terms. Anthropic's legal page says Team and Enterprise plans do not train on your content by default and that Enterprise admins set retention. Neither that page nor the plugin repository settles privilege or retention on your plan in contract terms, so read your agreement, check your regulator's guidance and ask Anthropic in writing first.
Which plugin should a very small firm try first?
Pick by the matter type you handle most: commercial agreements, employment or corporate work. Install one plugin, run its setup interview, and trial it on five closed or anonymised matters before using it on live work.
Further reads
- What an AI Consultant Does for a Law Firm, and What It Costs — What an AI consultant does for a law firm.
- ChatGPT or a Legal AI Tool: Which Should a Small Firm Use? — When a legal-specific tool beats a general chatbot.
- ChatGPT for Solicitors: Five Safe Tasks and Three to Avoid — Five safe tasks and three to avoid for a firm.
- AI Software for Solicitors: Data, Supervision and Supplier Checks — Data, supervision and supplier checks before buying.
- Claude v ChatGPT for Lawyers: A Same-Task Test on Your Own Documents — The same legal task run through both assistants.
- Copilot for Legal: What It Adds and Whether You Need a Paid Licence — Copilot for Legal adds contract review, drafting and chronology skills to Word and Cowork. Which licence it needs and what a small firm gets for the money.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Anthropic's Claude for Legal page and pricing page; Anthropic's claude-for-legal repository on GitHub (plugin list, install notes, privilege and review guidance); Anthropic's 12 May 2026 post 'Claude for the legal industry' on claude.com; Claude help centre articles on plugins and connectors.