Part 1 covered the research and the plan; Part 2 covered the RAG core — retrieval, generation, and citations that can't be hallucinated, all running locally on the M4 MacBook Air. This post is Phases 4 through 6: giving nsw-legal-research-assistant a user interface, exposing the same skills to Claude Code over MCP (and other MCP clients), and then running it live against a demo script taken directly from the interview where Thomson Reuters' Valerie McConnell walked Legal IT Insider through CoCounsel Legal's "Search a Database" and "Review Documents" skills — the McConnell interview documented back in Part 1's research.

Beat 1 — Grounded Q&A, scoped to a database

McConnell's demo queried named databases. That's the reason this app has a database selector. Asking a question with "All Documents" selected searches though every chunk of every uploaded document; scoping to one database (Rental Eviction Case Law (Jade)) will provide much more useful results.

During the COVID-19 pandemic, was there a moratorium on evicting tenants for rent arrears in NSW, and how did the court treat it in this case?

Example output:

A moratorium on evicting tenants for rent arrears existed during the COVID-19 pandemic in NSW. From [2], we see that the Residential Tenancies Regulation 2019 (NSW) was amended to include a new Part 6A "Response to COVID-19 pandemic," which provided for a "moratorium period" of six months (reg 41A). The moratorium did not apply in this specific case because the tenant was not an "impacted tenant" — there was no evidence to allow a finding that Mr Herbert was an "impacted tenant," and therefore the protections given by the regulation were not engaged.

Every claim carries a bracket marker. That's the whole trust mechanism in one screen — click one.

Beat 2 — Pinpoint citations, and the state where they can't be trusted blindly

Click [1] and the source PDF opens, scrolled to the page the details came from — the same "show its work" pattern McConnell described as CoCounsel's core trust mechanism, a hyperlink plus an excerpt backing every claim so a professional can verify without redoing the research. The citation metadata — case name, page number — is read directly off the retrieved database row, never generated by the model, so which excerpt gets cited is the LLM's call, but the citation itself can't be hallucinated.

Every citation in the run above is a verified source: a human found, downloaded, and uploaded the document, so its provenance is known — which is why those markers render as plain bold text with no warning.

Beat 3 — Review: one row per document

The Review skill asks a series of questions to every document individually.

Beat 4 — Skill-chaining: draft from an answer, no re-supply

McConnell's most complex demo move: approve a database answer, then say "draft a letter from that," with no re-upload and no re-running the search. Same move here — enter a description then click Draft:

draft a letter to the client summarizing this outcome

Real output:

Dear Client,

Re: COVID-19 Moratorium on Evictions in NSW

As you may recall, we were discussing the application of the COVID-19 moratorium on evicting tenants for rent arrears in NSW. In light of our previous research and findings [1], I am writing to inform you that a moratorium did exist during the pandemic period.

During this time, a new Part 6A "Response to COVID-19 pandemic" was added to the Residential Tenancies Regulation 2019 (NSW), which included a moratorium period of six months [2]. However, in our specific case, the court found that the protections given by this regulation were not engaged because you were not an "impacted tenant," as there was no evidence to support this designation [6].

In particular, the court noted that the tenant (you) was not a member of a household impacted by the pandemic, and therefore the moratorium did not apply in our specific case. As such, the court allowed the eviction proceedings to continue.

No new retrieval happens here — the exact answer and citations already on screen get passed straight back to a second LLM call, explicitly forbidden from introducing new facts or citations. It's deliberately stateless: the frontend, not the server, holds the "conversation" by re-sending what it already has, rather than a session table on the backend. That statelessness is also why the guardrail is worth testing honestly rather than cherry-picked: asking it to draft from an answer that hadn't actually established an outcome produced a correct refusal instead of a fabricated finding — a clean refusal on ungrounded ground is a stronger trust signal than a draft that always complies.

Beat 5 — Beyond the original demo

Not part of McConnell's demo — Model Context Protocol (MCP) wasn't discussed at all in the interview. ask_nsw_caselaw, introduced above, is the MCP tool matching Beat 1's Q&A skill. This same build shipped two more, mirroring the two skills it added: review_nsw_caselaw_collection for Beat 3, and draft_from_nsw_caselaw_answer for Beat 4. All three are thin wrappers calling the exact same functions the REST API and the frontend call — same skill, multiple interfaces.

A real MCP client/server round trip — spawning app/mcp_server.py as an actual subprocess over stdio, the same transport Claude Code uses, not a direct Python call — against the live stack:

Same shape as the browser — the same mix of grounded and not-found rows across the five documents, the same reused-figures-only drafting behavior — because both interfaces are calling review_collection() and draft_from_answer() directly, not a second reimplementation of them. The exact wording differs from Beats 3 and 4's screenshots, same as it does between any two runs of this app; the guarantee is architectural, not word-for-word reproducibility.

Where this actually runs

Every beat above ran against the full stack live on a Docker Compose setup on the M4 MacBook Air from Part 2 — Postgres/pgvector, the FastAPI backend, and the React frontend, all local, no cloud dependency for any of it:

Honest comparison to the original demo

What this app isn't trying to be: a reproduction of CoCounsel at enterprise scale. What it is: the same architectural pattern — curated, scoped RAG with a non-negotiable citation-grounding trust mechanism — proven against a real matter, with the actual gaps between "toy demo" and "useful tool" identified, closed where reasonable, and reported honestly where not.