Bring the framework into Claude or ChatGPT

Your AI gains the framework’s tools and its practice library. When you ask it for a piece of legal work, it works through the same five steps this site does: it asks what it needs to know, researches the rules and the law with sources, scores whether it should be used for this task at all, keeps to what the plan allows at each stage, and gives you a work record to check, approve and file.

Not for the framework. The public connector (/mcp) is free and needs no sign-in on this site. To give your AI your own firm’s policies, knowledge and the full precedent texts, add the signed-in connector (/mcp/firm) instead: it asks you to log in here once. Your AI plan decides whether you can add a connector at all: Claude allows custom connectors on Free (one connector), Pro, Max, Team and Enterprise plans, and ChatGPT offers them through developer mode on Business, Enterprise and Edu workspaces.

Both carry the framework’s tools and its practice library. /mcp is public: no sign-in, nothing about you is known, and precedents show their summary and AI-use guidance only. /mcp/firm needs you to log in on this site and approve the app once; it then also reads your firm’s approved AI policies, its knowledge base and the full text of the precedent library, as you, and can save drafts if you allow it. Use one or the other, not both.

Yes. Claude shows a caution whenever a connector is suggested by a link on a website, so you can check that the name and address come from a developer you trust: they should read AI Law Framework and https://ailawframework.com/mcp. It also notes that a connector without sign-in can be used by anyone who has its address. That is by design for /mcp: it is free, keeps nothing about your task and returns only the framework’s own published material. Leave Authentication on No sign-in, add no headers, and select Add. The firm connector (/mcp/firm) is different: Claude sends you to this site to log in and approve it.

In Claude, an Owner adds it once for the organisation, and each member then connects to it and switches it on. In ChatGPT, a workspace admin enables developer mode, creates the app and publishes it to the workspace. The organisation link on this page opens Claude’s dialog for an Owner.

Only by the minute it takes to answer the questions a careful supervising lawyer would ask anyway — and it picks up whatever you already said in your request. “Not sure” is always an acceptable answer; it simply raises the level of human involvement by one step, which is what doubt should do.

No. Each step needs the signed record the previous step returned, and the server refuses a record that has been altered. An answer the AI assumed instead of asking is refused too. What the server cannot do is check that the AI reported your words faithfully, so the work record shows where every answer and finding came from, for the lawyer to see.

It sees what your AI sends to its tools: the task as described, your answers and the findings with their sources. It uses them to apply the rules and keeps nothing about the task — no database row, no copy of the record — on either connector. Describe the task without client names where you can; the framework never needs them. The public connector (/mcp) keeps nothing at all. The firm connector (/mcp/firm) also keeps what a sign-in needs: which app you approved and when (your connection), the connection’s keys in a form that cannot be used if copied (only a one-way hash), and, for your firm’s owners and administrators, an access log of which app read which of your firm’s items and when. The log holds item ids, never your questions or any text.

The responsible lawyer. The AI proposes its own scores against the framework’s fixed wording, the evidence it found caps the governance score, and the record says plainly that they are a proposal for the lawyer to confirm or change.

The AI tells you so, with the reasons, and does not produce the output. Doing the work by hand is always one of the options and is scored beside the AI, so you still leave with a plan and a record — and often the reason is something you can fix, such as confirming a supplier term.

Yes. The method is the same anywhere: the AI researches the rules of the court, tribunal or regulator that will receive the work. The library is richest for Australian jurisdictions and also carries New Zealand and United States versions of many tasks.

Neither. It is a method for deciding how AI may be used on a task and for recording the human checks. It does not certify that a tool is safe, that the work is correct or that a saving is real; every judgement in the record is one a person makes or confirms.

So whoever reads them — a partner, a client, a court — knows which method was applied and can look it up. Every result the tools return carries the name and a link to this site, and the work record ends with it.

Each connection reads as the person who approved it, so it sees exactly what that person may see. Your firm’s owners and administrators can see the access log (who, which app, which tool, which items, when) and can switch AI agents off for the whole firm, which disconnects every app at once. Anyone can disconnect their own apps at Firm › AI agents.

In Claude or ChatGPT, remove the connector or app from the same settings page you added it on. In a terminal, run “claude mcp remove ailawframework” or “codex mcp remove ailawframework” (or ailawframework-firm). For the firm connector, also select Disconnect at Firm › AI agents: that cancels its access at once, even on a device you no longer have.