Claude for Legal

Claude for Legal is Anthropic’s open-source suite of legal agents and skills, organized by practice area, across commercial, corporate, employment, privacy, regulatory, AI governance, IP, and litigation work, among others. Anthropic publishes it under the Apache 2.0 license at github.com/anthropics/claude-for-legal. The agents are Anthropic’s work, and the credit for them is Anthropic’s; Sonata uses that published library, it does not author it.

Sonata imports a copy of the Claude for Legal agents, curates which of them fit a chat-with-an-agent surface, and runs them inside its own governed platform, under your brand, your access controls, and the policies you set. Your team gets a vetted practice-area library that behaves like every other agent in Sonata, rather than going elsewhere, installing plugins, and managing that themselves. The library is something your organization chooses to show and curate; it is brought in deliberately and kept under your governance, not wired to update on its own.

On each department launchpad your agents are grouped, and the groups have different origins worth keeping straight. Approved agents are your department’s own, vetted by your team. Powered by Sonata are agents Sonata builds and ships itself: first-party, general-purpose tools, fully managed so they cannot be edited, and yours to copy into an editable version. These are not Claude for Legal; they are Sonata’s own. Claude for Legal is the curated library described here, Anthropic’s agents governed by Sonata. My agents are the ones your own people create. The workspace guide walks through all of the groups; this guide is about the Claude for Legal one.

Running Claude for Legal inside Sonata, rather than alongside it, means the same controls apply to it as to everything else.

The agents stay as vetted. A Claude for Legal agent’s wording, its name, description, and instructions, is managed and shown read-only, so it stays the version it was vetted as. An administrator can still adjust the operational settings that are yours to set, like which model it runs on and the references it draws from, and anyone can copy one into a fully editable My agent. The Powered by Sonata agents are stricter still: fully locked, copy to change.

The trustworthy work is code, not the model. Sonata’s most exact operations, like Document Comparison and Structured Query, do not ask the model to be careful and hope. The comparison or the count is computed in plain code first, and that result is handed to the model as something it must explain and cannot change. The answer you can check is produced by deterministic code; the model puts it into words. This is the strongest sense in which Sonata keeps control: the part that has to be right is not left to a model’s judgment.

Access is governed, per organization. Who can use what is enforced by roles and at the database, not just in the interface: people see only the departments they are given, only administrators reach the governance surfaces, and Policy and access holds the organization’s levers in one place, including the ceiling on what a connected system an agent uses may do. Claude for Legal agents live inside that same boundary.

The model and its cost are yours. You set the default model your agents run on, and you choose whether they run on the Sonata managed key or your organization’s own provider key. With your own key, inference runs under your own provider agreement and your own cost, for Claude for Legal agents the same as any other.

Whether the Claude for Legal library shows at all is your decision, made in Policy and access under Connections, in the Content area where the organization’s curated content libraries are listed. Turn it off and its agents stop appearing across the departments; turn it on and they return. The Connections guide covers that control in full.

Claude for Legal agents run on the same Anthropic API as the rest of Sonata, and under Anthropic’s standard API terms that means your inputs are not used to train the model. That is a property of the API being called, not a switch Sonata adds. As with any agent, when one proposes an action that would change something outside Sonata, like sending a message through a connected system, that specific write is designed to wait for a person to approve it; reading and reasoning are not gated this way.

Claude for Legal is used under the Apache 2.0 license (apache.org/licenses/LICENSE-2.0), and the credit for the agents belongs to Anthropic. This page records that provenance and credit; it is not itself the project’s formal license notice.