What do these terms mean?
| Term | Plain-language explanation | Question to ask |
|---|---|---|
| AI assistant | A system that responds to instructions and information | What evidence supports its answer? |
| Plugin or skill | A set of capabilities or structured instructions for a task | What task is it designed to perform? |
| Agent | A workflow that may perform several linked steps | Who reviews each important step? |
| Connector | A bridge to another application or source of records | Can it only read, or can it also change data? |
A product may combine several of these functions. Brand names alone do not reveal the full permission model or current feature availability.
Browse documented examples by task
Anthropic publishes a legal connector and plugin marketplace and an open legal-plugin repository. Those are useful official starting points for evaluating the following categories.
| Legal work | Example of a documented direction | What remains for the user to verify |
|---|---|---|
| Finding cases and legal material | CourtListener and legal research connections | Coverage, citation checks, access rights |
| Working with contract records | Ironclad or Docusign connections | Authorized records, modification abilities |
| Reviewing matter files | iManage or Everlaw-related connections | Document access and privileged information |
| Organizing legal tasks | Practice-specific plugin workflows | Actual scope and required professional supervision |
| Supporting litigation preparation | Research and evidence-oriented workflows | Accuracy, source lineage and court obligations |
These examples are vendor-documented options, not test results or a list of recommended suppliers. The repository itself explains that lawyers must review and take responsibility for work product; plugin suggestions do not replace professional analysis.
How should a legal team supervise an agent?
Begin with a workflow map: input, sources consulted, intermediate steps, proposed output and final approval. A research-only task may permit the tool to retrieve public information, while a case-management task may carry additional consequences if it alters a deadline or record.
Ask: When is human review mandatory? How are sources shown? Which action creates an external effect? Can access be revoked? What happens when the system cannot verify an answer? A useful agent should be evaluated not only on successful demonstrations but also on incorrect or incomplete outputs.
Understand data access and software connections
Some tools work only with information pasted into a conversation. Others retrieve documents from connected applications. Anthropic's connector guidance describes how service permissions shape access and how organizations can restrict certain actions.
Before activating a connector, test access using approved sample records and confirm that users cannot access material outside their role. If the tool can change records, require a separate action review and a clear log.
Read-only versus action-taking features
A read-only task might retrieve a published case or locate a permitted contract clause. An action-taking task might create a new matter note or submit a change to a contract system. Both can create errors; action-taking tools also need closer attention to authorization, reversibility and audit trails.
An agent that presents a completed workflow without showing the material it relied on may be difficult to check. Ask the vendor how evidence and the sequence of actions can be inspected.
A practical plugin review card
For each listed product or feature, record the vendor, official feature link, legal task, source-data access, permissions, output type, required approvals, available logging, last-verified date and unresolved restrictions. Mark unknown features as unknown; do not infer that an integration shown in a demonstration is available in every account.
Sources and editorial independence
Relevant primary sources include the Anthropic legal marketplace, its legal-workflow repository and connector help center. Vendor descriptions help establish what is advertised, not that a solution is accurate, independent or suitable for a particular legal matter. No partnership with any named vendor is represented.
Frequently asked questions
Can a legal AI agent make a legal decision for us? A tool may support analysis or workflow steps, but the responsible professional must retain the appropriate judgment, supervision and approval duties.
Is a plugin safe because it comes from a marketplace? Do not assume so. Examine the source, permissions, data handling, actual functionality and your organization's requirements.
Do we need a connector for every agent? No. Some workflows use only the information supplied to them; others require access to outside systems.
Are agents and legal AI integrations the same directory topic? No. An integration concerns how systems connect. This page explains the tasks and behavior of assistants, agents and plugins.
Explore documented agents and plugins
Start with one task that can be observed and checked. Read the official vendor material, determine permissions, and test the workflow without confidential information until approvals are in place. Visit Legal AI Integrations for connection-specific questions (B02-001).
Find your next useful step
Start with the sources and practical checks above, then continue to a related guide.
Browse Agents and Plugins