Claude for Legal: Anthropic's Plugin Suite for Legal Workflows

What claude-for-legal actually is, and who it is built for

The repository is a reference collection, not an application. It holds practice-area plugins (commercial, corporate, employment, privacy, product, litigation, regulatory, AI governance, IP), managed-agent cookbooks for scheduled workflows, MCP connectors, and a set of named agents that each map to one workflow. The README frames the audience as in-house commercial, privacy, product, corporate, employment, litigation, regulatory, AI governance and IP teams, plus the learning side of the practice: law school clinics and students. Two directories, legal-clinic and law-student, exist for that second audience.

The important framing sits in a callout near the top of the README. Every output is described as a draft for attorney review, not legal advice, not a legal conclusion, and not a substitute for a lawyer. The same callout states the plugins do not represent Anthropic's legal positions. If you are evaluating this as a way to remove lawyer time from a workflow, you are evaluating the wrong thing. The design assumption is that a lawyer reviews, verifies and takes professional responsibility for anything leaving the building.

One source, two deployment targets, and the practice profile that ties them together

The README's central architectural claim is that everything is available two ways from one source: as a Claude Cowork or Claude Code plugin, or deployed through the Claude Managed Agents API behind your own workflow engine. The same system prompt and the same skills are used in both cases. What changes is where the agent runs and what triggers it.

The mechanism that makes this portable is the practice profile. Each practice-area plugin is built around two pieces: a cold-start interview that learns your playbook, and a CLAUDE.md practice profile that every skill reads from. Skills do not carry your house positions themselves; they read them from the profile. This is why the README tells you to start with the agents matching your work and then tune the underlying skill, the practice profile and the connectors.

Agents divide into two execution shapes. Interactive ones run as slash commands, for example /commercial-legal:review for the Vendor Agreement Reviewer and the NDA Triager, or /corporate-legal:tabular-review for Tabular Diligence Review. The second shape is the scheduled agent, listed in the command column as such for Renewal Watcher, Deal Debrief, Playbook Monitor, Data Room Watcher and Leave Tracker. Those are the ones the managed-agent cookbooks exist to support, because a job that scans a contract register for cancel-by deadlines has to run on a schedule rather than when someone types a command.

Connectors are the third layer and the one most likely to decide whether a given workflow works for you. The README lists general productivity systems (Slack, Google Drive, Box) and legal-specific ones (Ironclad, DocuSign, iManage, Everlaw, CourtListener, among others). A docket watcher without CourtListener access, or a contract register scan without your CLM, is a different proposition from the demo.

Installing the plugin and running your first review

The README points new users at QUICKSTART.md, which it says covers installation in 60 seconds, and describes the README itself as the full reference. For Cowork the stated path is to install Claude Desktop, get access to Claude Cowork, and follow the video linked in the README. The reliable instruction is to follow QUICKSTART.md.

Once the suite is installed, the interactive entry points are slash commands scoped to a plugin. The Vendor Agreement Reviewer and the NDA Triager both use the same command, which is worth noting: the triage and the full review are the same surface, and what distinguishes them is the input and the playbook the skill reads.

/commercial-legal:review

Run against an inbound NDA, the README says the expected result is a GREEN/YELLOW/RED triage so that only the hard ones reach a lawyer's desk. Run against a vendor MSA, the same command produces a redline memo against your playbook. If you have not yet completed the cold-start interview, the skill has no playbook to compare against, so do that first.

The corporate side has a different shape. Tabular Diligence Review produces one row per document with every cell cited, which is the property to check on your first run: an uncited cell means the source attribution guardrail did not fire.

/corporate-legal:tabular-review

For a scheduled workflow, the repository ships managed-agent cookbooks rather than a slash command. The README names five: renewal watcher, docket watcher, regulatory feed monitor, diligence grid and launch radar. Deploying one means running it through the Managed Agents API behind your own workflow engine, which also means you own the scheduling, the credentials and the failure handling.

Where the design breaks down

The guardrails are conservative by design, and that is a cost as well as a benefit. The README lists source attribution on every citation, conservative defaults on privilege and subjective legal calls, jurisdiction assumptions surfaced, and explicit gates before anything is filed, sent or relied on. Each of those is a checkpoint a human has to clear. A team expecting the NDA Triager to clear a queue without a lawyer looking at the YELLOW and RED items has misread the tool.

The jurisdiction handling is the sharper limitation. Employment workflows such as Termination Reviewer, Hire Reviewer, Worker Classification Screener and Policy Drafter are explicitly jurisdiction-specific, and Policy Drafter is described as producing state supplements where law differs. Worker Classification Screener tests an engagement against the controlling state test. These are US-shaped mechanisms. The international surface in the agent table is International Expansion Planner, which kicks off EOR-versus-entity planning and outside-counsel briefing rather than answering the substantive question.

The README also states that the law in many of these areas is unsettled and evolving. That is not boilerplate; it means a skill's risk flag or characterization of case law is a starting point for the reviewing attorney's own analysis. Where a skill offers a checklist item or a suggested framework, the README is explicit that this is an aid, not Anthropic's view of the law.

Finally, the whole thing is Claude-shaped. The deployment options are Cowork, Claude Code and the Managed Agents API.

How this differs from a legal research platform

Compared with a hosted legal research and drafting platform, the difference is structural rather than a matter of feature counts. Such a platform is a hosted product with its own interface, its own document store and its own model pipeline; you buy access and your lawyers work inside it. claude-for-legal is the opposite arrangement. It is a set of prompts, skills and connector definitions that you install into a general-purpose agent runtime you already have, and you bring the document systems through MCP connectors such as iManage, Ironclad or Box.

That difference cuts both ways. You get the playbook in a CLAUDE.md file you can read and edit, and the same skill can run interactively or on a schedule through the Managed Agents API. You also own the integration work, the connector credentials and the scheduling. A team without an engineer or a platform owner to wire up the connectors will find the hosted alternative faster to stand up, and a team that wants its review standards version-controlled in its own repository will find this arrangement easier to audit.

The licence is Apache-2.0, which permits commercial use and modification subject to the terms of that licence. The repository also carries a CLA.md and a CONTRIBUTING.md, so contributions run through a contributor licence agreement. Nothing here is legal advice about your own compliance obligations; read the licence text and the CLA before you fork or redistribute.

Maintenance, upgrades and what you are taking on

The last push to the default branch was on 2026-10-29, and the repository is not archived. The repository has no published GitHub releases, so there is no versioned artefact to pin against. Upgrades therefore mean tracking the main branch, and because skills read from a CLAUDE.md practice profile that you edit, a change to a skill's expectations can collide with your local profile. Keep your practice profiles in your own version control separately from the upstream tree so you can diff them after a pull.

There is no changelog among the repository's top-level files. If you deploy a scheduled agent through the Managed Agents API, the upgrade path for the cookbook is yours to manage, and so is the question of what happens to a renewal watcher run that fails halfway through a contract register scan.

The licence position is straightforward in outline: Apache-2.0 on the code, a CLA for contributions. What that does not settle is the professional responsibility question, which the README answers directly. The attorney using the plugin, not the plugin and not Anthropic, is responsible for the legal positions taken in the work product.

Editorial conclusion

Adopt it if you already run Claude Code or Claude Cowork and want a starting point for contract review, diligence or employment workflows that a lawyer will review anyway. Do not adopt it as a substitute for legal judgement or as a drop-in replacement for a legal research platform; the README states plainly that outputs are drafts, not advice. Before you build on it, verify the install path in QUICKSTART.md, check which connectors your team actually has, and confirm that the jurisdiction assumptions each skill surfaces match the jurisdictions you practise in.

Frequently asked questions

Official sources

1. anthropics/claude-for-legal on GitHub
2. Issues
3. License: Apache-2.0
4. README