Introducing Clio for Codex

Today we’re making Vincent available in Codex as a plugin. The premise is simple: when you’re doing legal work on Codex, you should be able to call on Vincent as a legal agent. 

Through MCP, you can now invoke Vincent directly from Codex. This means you can move from analyzing a docket, to researching a judge, finding the controlling law, developing a strategy, and producing finished work without rebuilding the legal context at every step. Vincent brings that context with it: legal data, court records, matter information, purpose-built legal workflows, and source provenance. MCP gives Codex a standardized way to call on those capabilities as your work evolves.

The plugin design

We had a fundamental choice to make when building this: we could connect Clio Library and Docket datasets and let Codex query them directly, or we could give you access to Vincent. We chose Vincent.

Access to authoritative data is only part of the story. How you search it, which sources you pursue, how you test an argument against the record, when you broaden or narrow the inquiry, and how you connect authorities back to the question all serve a purpose. Vincent already knows how to do this work extremely well.

When Codex invokes Vincent, it isn’t simply querying Clio’s data. It is calling a legal agent we have built, tested, and evaluated specifically for legal research and analysis. Vincent brings the research process with it: purpose-built legal skills, strategies for querying Clio Library and Docket, model and reasoning configurations selected against our evaluation sets, and the ability to delegate complex tasks to a multi-agent pipeline that can analyze large bodies of material in parallel.

That gives us something a data connector cannot: accountability for the legal work being returned to you.

We control the environment in which Vincent performs that work, from the quality of the underlying datasets to the agent harness, models, configuration parameters, research instructions, and evaluation framework. When we improve how Vincent researches a legal question, Codex gets the benefit of that improvement without having to learn the research strategy itself.

The architecture also gives each legal research task a fresh context window. Long agentic sessions accumulate context, and that context can degrade over time. Your legal research should not inherit context rot because it happens to be step twelve of a much larger task. Vincent gets the attention budget to do the legal work properly, then returns the result to Codex.

An agent-to-agent architecture carries the legal research environment we’ve built around it: the models, skills, instructions, research strategies, evaluation systems, and controls that determine how Vincent approaches legal work in the first place. In other words, you get Vincent as we’ve built it, powered by the data behind it.

Show the work

Trustworthy legal research starts with being able to inspect the work behind it.

The plugin uses MCP to surface Vincent’s progress, research trail, and exact output directly inside Codex. You can see what Vincent returned and distinguish it from what Codex subsequently produced, including work that combines Vincent’s output with other data sources or skills.

This is important visibility when multiple agents are contributing to your work. You should be able to understand where the legal analysis came from and inspect the authorities behind it.

Extending the work in Codex

Once Vincent finishes the legal task, you’re back in Codex with the result and the full capabilities of Codex still available to you.

You can combine Vincent’s analysis with other datasets, create your work product, write and execute code, or perform further analysis. We also give Codex tools to retrieve the underlying legal authorities directly, so you can continue working with the sources themselves rather than treating Vincent’s answer as the end of the process.

The boundary between the two agents is intentional. Vincent gets a dedicated context for the legal task, while Codex keeps the broader context of what you’re trying to accomplish. That concentrates each agent’s attention on the work it is best equipped to do, although it also means Vincent only sees the parts of the broader Codex conversation passed into its context.

It also keeps legal research inside Vincent’s environment, with the models, research skills, context, and compute we have built and evaluated for that work. Vincent can do the legal work in that controlled environment, then hand the result back to Codex with its broader capabilities and context intact.

The result is a deliberate division of labor: Vincent goes deep on the legal context, and Codex takes the work forward.

Engineering for regionality

You should not have to choose between a global protocol and the regional data architecture you rely on. 

Historically, we have delegated region awareness at the DNS layer. A client chooses its region before initiating a request: traffic to eu.api.clio.com can only be routed to systems in the EU, and traffic to ca.api.clio.com can only be routed to systems in Canada. That gives you a top level guarantee about where your request will be processed, and where the data associated with it resides. 

But MCP endpoints work differently, they are global by design. The specification does not provide a mechanism for the client to declare the region to which an individual request should be routed, or for us to require one. From Codex’s perspective, there is a single endpoint: mcp.api.clio.com. We wanted to preserve the simplicity of that global endpoint without giving up our regional architecture, so we built a global MCP broker. 

When you connect through Codex, the broker identifies the region associated with your Clio account and routes the request to the appropriate regional infrastructure. Codex can work through a single global endpoint while Clio handles regional routing behind the scenes.

You see one MCP endpoint, but your request still goes where it belongs. This lets you access Vincent through a global protocol while preserving the regional architecture underneath it. The same principle runs through the entire implementation. The point isn’t simply to make more Clio data available inside Codex, rather, it’s to bring the capabilities we’ve built around that data with it.

You can call on Vincent for the legal work, see the research and authorities behind it, and then keep going in Codex with the broader task still in hand.

Building for an agentic legal ecosystem

Legal work is increasingly happening across AI environments, and we want you to be able to access Clio’s legal AI wherever that work happens. MCP gives us a way to bring Vincent into those environments with the depth behind it intact. That includes the legal data, court records, research methods, matter context, provenance, and controls we’ve spent years building.

Making Vincent available in Codex, through a new MCP plugin is an important step in that direction.