Claude for Lawyers: What It Does Well and How to Connect It

Claude is good for lawyers at drafting, summarizing long documents and reviewing contracts against a playbook, as long as a lawyer verifies every citation it produces. Anthropic's Claude for Legal adds practice-area plugins on top of that. With a connector, Claude can also act on client-facing work, such as sending a document request, but the lawyer makes every call that matters.

Key Takeaways

• Claude is strong at reading long documents, drafting in a house style and structured contract review; it still invents citations, so nothing it names as authority goes into a filing unverified.
• Claude for Legal is Anthropic's own plugin and connector set for legal work: it drafts and reviews inside the firm's systems, and it does not collect documents from a client.
• The open dokutrak-mcp connector lets Claude send a client a document request and chase a rejected file from inside Claude Desktop or Claude Code. It cannot accept a document; that stays with the lawyer.
• ABA Opinion 512 requires the client's informed consent before case information goes into a self-learning tool, and reading a request's status through a connector counts as case information.

This page is written for the lawyer, not the client. It isn't legal or ethics advice, and your state bar's guidance governs over anything here. One disclosure up front: I build DokuTrak, a client document request tool, which is why the connector section below is the one I know best.

Is Claude good for lawyers?

Yes, for the reading-and-writing half of the job. Claude is strong at getting through a long document (a deposition transcript, a lease, a 40-page contract) and returning a usable summary or a first draft in something close to the firm's own voice. It is also useful for a structured review pass: checking a contract against a playbook of clauses the firm already uses, or triaging a stack of NDAs for the ones that deviate from the standard.

The one failure that matters is the same one every general-purpose model has: Claude has no citator and no authoritative case database, so it can produce a citation that does not exist and present it with complete confidence. Courts have sanctioned attorneys in order after order since 2023 over exactly this failure, regardless of which model was involved. The rule for Claude is the same as for any of them: verify every case, statute or rule it names before it reaches a filing, a brief or a client.

What is Claude for Legal?

Claude for Legal is Anthropic's set of legal plugins, agents and connectors for Claude. Anthropic publishes them in an open GitHub repository, installable as a Claude Cowork or Claude Code plugin, and describes practice-area plugins for commercial, corporate, IP and litigation work, with connectors to tools such as iManage, NetDocuments, Docusign, Ironclad and Thomson Reuters.¹

The repository groups its plugins by the legal workflows Anthropic says it sees most: in-house commercial, privacy, product, corporate, employment, litigation, regulatory, AI governance, IP, and the learning side of the practice (law school clinics and students).² It names specific agents inside those plugins, including an NDA Triager, a Vendor Agreement Reviewer, a Matter Intake agent and a Docket Watcher. Anthropic is explicit about the ceiling on all of it: "Every output from these plugins is a draft for attorney review — not legal advice."²

That ceiling is worth sitting with, because it draws the line this page is built around. The plugins work on contracts, filings, research and the firm's own document systems: reviewing, drafting, triaging. None of them sends a client a request for documents or chases a file that didn't show up. The closest thing is in the legal-clinic plugin, which includes a "Client Letter Drafter" described for "routine client correspondence — appointment confirms, doc requests, updates."² It drafts that letter. The README describes no step that sends it or tracks it.

Read Anthropic's own legal solutions page and the claude-for-legal repository before installing anything — they describe exactly what each plugin does and doesn't do.

What lawyers use Claude for: 10 tasks

The table below is the whole page in one place. The last column never moves: whoever signs, files or sends is the one who answers for it, whether Claude touched the draft or not.

Rows 1 through 7 are the work Claude for Legal's plugins are built for: drafting, reviewing, triaging, inside the firm's own systems. Rows 8 through 10 are a different kind of task, reaching a client who isn't inside those systems at all, and that's where a connector built for client-facing work does something the legal plugins don't.

Connectors: from your inbox to the client file

A connector lets Claude act in another system instead of only writing text. Anthropic lists legal connectors for document management, e-signature and research. For client-facing work, the open dokutrak-mcp connector lets Claude send a client a document request, chase the files that were rejected, and report where a request stands. It cannot accept or reject a document.

Where it runs today: Claude Desktop and Claude Code, as a local MCP server you run with npx -y dokutrak-mcp, authenticated with an API key you create from Settings → Connect an agent in DokuTrak. This local connector runs only where a local MCP server can: claude.ai on the web and the mobile apps connect to remote servers instead.³

Connecting it in Claude Code takes one command:

claude mcp add dokutrak -e DOKUTRAK_API_KEY=dk_live_… -- npx -y dokutrak-mcp

Or, in Claude Desktop's config file:

{
  "mcpServers": {
    "dokutrak": {
      "command": "npx",
      "args": ["-y", "dokutrak-mcp"],
      "env": { "DOKUTRAK_API_KEY": "dk_live_…" }
    }
  }
}

The connector exposes four tools. create_request creates a document request and sends it in one call, but nothing goes out without your yes: the documentation notes that "Claude Code prompts each time, even in auto or bypass mode, and Claude Desktop treats it as a tool that always needs approval."³ request_replacement moves a rejected file back onto the client's plate and back onto the reminder schedule; "This call sends no email by itself; the reminders do."³ get_request reports a request's status: its checklist, each file's verdict, and where the reminders stand. download_documents returns the request's files as one zip, embedded directly in the conversation.

What it refuses outright: accepting or rejecting a document ("Approving or rejecting a document is your decision, taken in the DokuTrak dashboard. No tool here can take it"³), along with billing, workspace settings and API keys. Revoking the connection takes effect on the very next call.

Here's what that looks like as a conversation. This is an illustrative example, not a real client or matter:

You: Send Maria Lopez the intake request: photo ID, insurance card, the police report, and medical bills so far. Deadline Friday.

Claude: Here's the draft — recipient, deadline, checklist and message. Send it?

You: Yes.

(later)

You: What's still missing on Lopez?

Claude: 3 of 4 received. The police report was rejected as unreadable. I can ask her for a replacement.

Read the public dokutrak-mcp repository and its npm page before connecting it to a real matter: they're the source for exactly what the server can and cannot do. And one thing worth saying plainly: you don't need Claude to use DokuTrak. The connector is a channel for firms that already work inside Claude; DokuTrak's web app sends and chases on its own either way, and its optional AI first pass flags files.

Confidentiality, client data and ABA Opinion 512

ABA Formal Opinion 512, issued July 29, 2024, calls for the client's informed consent before information about the representation goes into a self-learning AI tool — and a boilerplate waiver in an engagement letter doesn't meet that bar. ChatGPT for lawyers covers the opinion's competence, confidentiality, supervision and billing points in more detail; the short version, sourced to the same NCBE summary, is that consent has to name why the tool is used, the specific risk, and what kind of client information might be disclosed.⁴

A paid plan changes a different question than the consent one. Anthropic's pricing page states "No model training on your content by default" for Team, and shows "None by default" for Enterprise.⁵ That answers whether Claude's maker trains on what you type. It says nothing about whether your client agreed to have their matter discussed with an AI tool in the first place. That's Opinion 512's question, and a plan's training default doesn't answer it.

The connector adds a version of this that's easy to miss because nothing gets typed into a chat box. Asking Claude for a request's status sends that client's name, email and file verdicts into the conversation. Downloading the files puts the client's actual documents into it. Treat both the same way you'd treat pasting the documents directly: the same consent analysis from Opinion 512 applies, whether the client information arrived by typing or by a connector call.

Claude vs ChatGPT for lawyers

Neither one wins on brand. What actually changes a firm's exposure is the plan it's on and the policy it writes, because data-handling defaults differ between consumer and business tiers. Both models invent citations with the same confidence, so neither is a substitute for a citator-backed research tool. ChatGPT for lawyers runs the comparable task table for ChatGPT, including the one task, chasing a client for documents, that neither general assistant can do without a connector built for it.

The real difference worth naming is Claude for Legal's plugin set: practice-area agents installed directly into Claude Cowork or Claude Code, reviewing and drafting against the firm's own playbooks. Whether that's worth switching for depends on which plan your firm already pays for and trusts with client text, not which assistant writes a cleaner paragraph.

How much is Claude for lawyers?

As of October 2026, Anthropic's pricing page lists Pro at $17 per month billed annually ($20 if billed monthly). Team has two seat types: Standard at $20 per seat/month billed annually ($25 monthly), and Premium at $100 per seat/month billed annually ($125 monthly). Enterprise is listed as "US$20/seat/month, billed annually," plus usage billed at API rates.⁵

The confidentiality statement that matters for client work (no training on your content by default) attaches to Team and Enterprise, not to Pro.⁵ A solo practitioner on Pro is on the opt-out default; a firm handling client matters on Claude wants Team or above. Figures change; check claude.com/pricing for the current numbers before you buy.

Which version of Claude is best for legal work?

The plan matters more than the specific model for most firms. Anthropic's model line changes every few months, and a model name you read about today may not be the one your plan offers tomorrow. The practical rule: use the current top model your plan makes available for long-document review, and verify its output the same way regardless of which one it is. The citation risk doesn't change with the model.

What Claude should never do alone in a law firm

• Advise a client. Draft the explanation, then have the lawyer deliver it.
• File or send anything without review. The connector's confirmation step exists for exactly this reason: nothing reaches a client without your yes first.
• Accept a client's document. That call stays in the DokuTrak dashboard, never in a chat window.
• Settle a privilege question. None of the vendor pages cited here says what happens to privilege when client information enters its tool. Ask your bar.
• Bill for it. Time the tool saved you is not time you worked; only actual hours get billed.

The AI flags what needs attention. The lawyer makes the call on every one of them.

---

Start a 14-day free trial, no card. Add one only to keep going. Solo, Team and Agency plans, detailed on the pricing page linked above.

For the document chase itself rather than the AI layer on top of it, see ChatGPT for lawyers, AI tools for small law firms, and how DokuTrak's AI first-pass check flags a wrong or unreadable file before you open it. For the accounting side of the same connector, see Claude for accountants. The law firms page and DokuTrak features cover the rest of what the product does beyond Claude.

Sources

Footnotes

1. Anthropic, Claude Legal Solutions, retrieved October 2026: "Claude connects to iManage, NetDocuments, Docusign, Ironclad, and Thomson Reuters." · "New plugins for commercial legal, corporate legal, IP, litigation and more — configured for your practice area." https://claude.com/solutions/legal ↩

2. Anthropic, anthropics/claude-for-legal README, retrieved October 2026: "Reference agents, skills, and data connectors for the legal workflows we see most — in-house commercial, privacy, product, corporate, employment, litigation, regulatory, AI governance, IP, and the learning side of the practice (law school clinics and students)." · "Every output from these plugins is a draft for attorney review — not legal advice." · Client Letter Drafter: "Routine client correspondence — appointment confirms, doc requests, updates." https://github.com/anthropics/claude-for-legal ↩ ↩² ↩³

3. dokutrak-mcp README, v0.1.2, retrieved October 2026: "Nothing goes out without your yes… Claude Code prompts each time, even in auto or bypass mode, and Claude Desktop treats it as a tool that always needs approval." · "This call sends no email by itself; the reminders do." · "Approving or rejecting a document is your decision, taken in the DokuTrak dashboard. No tool here can take it." https://github.com/Crackx17/dokutrak-mcp ↩ ↩² ↩³ ↩⁴

4. NCBE, The Bar Examiner, "Generative Artificial Intelligence Tools," Fall 2024, summarizing ABA Formal Opinion 512 (July 29, 2024). Retrieved October 2026. https://thebarexaminer.ncbex.org/article/fall-2024/generative-artificial-intelligence-tools/ ↩

5. Anthropic, Pricing, retrieved October 2026: Pro $17/month annual, $20/month monthly; Team Standard seat $20/seat/month annual, $25/seat/month monthly; Team Premium seat $100/seat/month annual, $125/seat/month monthly; Enterprise "US$20/seat/month, billed annually" plus usage at API rates. Model training: "No model training on your content by default" (Team), "None by default" (Enterprise). https://claude.com/pricing ↩ ↩² ↩³