Your Case File Is Stuck Outside Claude — Here's Why That Matters

If you've ever typed "what did the officer actually say at 14:32?" into Claude and gotten a confident, useless answer, you already know the problem. Claude drafts a fine suppression motion and summarizes case law brilliantly — but it has never seen your bodycam footage, your client's jail calls, or the police report you're trying to impeach. The model that does your best writing knows nothing about your case.

The workaround most defenders try is pasting discovery into the chat. That fails twice. It doesn't scale — a single case can be 40 hours of video, hundreds of pages of scanned reports, and audio files Claude can't ingest at all. And it's a data-handling decision you're making on the fly, with client confidences, inside a general-purpose chat window.

This isn't a prompt problem. It's a connection problem — and the industry has a name for it. Without a standard, every AI application needs a custom integration for every data source: ten AI apps times a hundred tools means a thousand bespoke connectors. Anthropic's Model Context Protocol, introduced in November 2024, solves this the way USB-C did for hardware — one standard port, so any compliant client can talk to any compliant server. It's why people call MCP "USB-C for AI applications".

The pattern is already live in legal work. Thomson Reuters and Free Law Project both launched MCP integrations connecting Claude to live legal research — Westlaw's corpus on one end, free CourtListener dockets on the other. Notably, the most common MCP use in the wild is exactly what defenders need: read-oriented access to data sources and knowledge bases, at 63% of connected server types per Zuplo's State of MCP report. Research is connected. Your evidence isn't — yet.

What a read-only evidence connection looks like in practice:

• List your cases and the evidence in each, without Claude ever writing to your files
• Search across transcripts and OCR'd documents with word-level timestamps
• Read a window of a transcript around a specific moment — like 14:32 on the bodycam
• Pull the citation for an answer back to the exact second of video or page of the report

That's the shape of Evidence Copilot's MCP server: five read-only functions — list_cases, list_evidence, search_evidence, read_window, get_citation — exposed to Claude Desktop and Claude Code, scoped to your organization, with revocable tokens and full audit logging. No defense-side competitor offers an MCP integration, which tells you how early this is.

The security design matters as much as the capability. Security is the number-one blocker to MCP adoption — 50% of builders cite access-control complexity, and a Fudan University study found 40.55% of live remote MCP servers exposed tools with no authentication at all. A read-only, org-scoped, audited server is the direct answer to that concern: Claude can look, but it cannot touch your originals.

The gap between "Claude writes my motions" and "Claude knows my case" is the last mile of AI-assisted defense work. MCP is how you close it.

What an MCP Server Actually Does (and Why Read-Only Is the Point)

Ask Claude a question and it answers from its training data. Connect an MCP server, and it answers from your data — live, in the moment. That distinction matters when the question is "what did the officer say at minute 14 of the bodycam," not trivia.

MCP (Model Context Protocol) is an open standard Anthropic introduced in November 2024 that lets AI assistants connect to external data sources and tools in real time. Anthropic's announcement framed it as a way for developers to build against one standard protocol instead of maintaining a separate connector for every data source. The a16z analysis calls it "USB-C for AI applications" — plug in a server, and any MCP client gains new capabilities.

Under the hood, an MCP server exposes three primitives:

• Tools — functions the AI can invoke, like running a search or pulling a record.
• Resources — data the AI can read, like documents or database entries.
• Prompts — pre-defined templates that shape how the AI approaches a task.

Here's the part that matters for practitioners: the dominant real-world use of MCP is read-oriented data access. In a State of MCP survey, 63% of users connect servers that are data sources — documentation, knowledge bases — and the top-cited value, at 30%, is better context for AI responses. Not automation. Not agents taking actions. Better answers, grounded in real data.

Legal is already there. On the same day, Thomson Reuters and Free Law Project launched Claude MCP integrations — CoCounsel exposing a 1.9-billion-document Westlaw corpus, CourtListener exposing free public court decisions, PACER data, and oral argument audio. Anthropic simultaneously released 20+ connectors to legal tech products. The pattern is mainstream: point Claude at an authoritative legal database, and it answers from that database instead of memory.

But notice what those servers cover: research. Case law, dockets, citations. None of them connect Claude to your case evidence — the discovery dump sitting on your desk. The bodycam, the interrogation footage, the jail calls, the police report with the timeline that doesn't quite match the video.

That's the gap Evidence Copilot's MCP server fills. It exposes five read-only tools to Claude Desktop and Claude Code — list_cases, list_evidence, search_evidence, read_window, and get_citation — so you can ask Claude what's in a case and get answers drawn from the actual evidence, with citations back to the exact source. Read-only is the point: the server can't modify, delete, or move anything. It's org-scoped, uses revocable tokens, and every query is audited — which matters in a landscape where a Fudan University study found 40.55% of live remote MCP servers expose tools with no authentication.

Research MCPs tell Claude what the law says. An evidence MCP tells it what your case actually contains.

The Five Tools: How Evidence Copilot Exposes Your Case to Claude

When you connect Evidence Copilot's MCP server to Claude Desktop or Claude Code, you get exactly five tools — and their names tell you almost everything about the design philosophy. They are list_cases, list_evidence, search_evidence, read_window, and get_citation. All five are read-only, org-scoped, revocable, and fully audited. Claude can look at your case file; it cannot change it.

That read-only posture is deliberate. The most common MCP server type by far is data sources and knowledge bases at 63%, per a state-of-MCP survey, and security remains the top adoption blocker — 50% of builders cite access-control complexity. With roughly 40% of live remote MCP servers exposing tools with no authentication, according to a security analysis of MCP servers, a defense-side integration has to be conservative by default.

Here's what the workflow looks like in practice, say, prepping the stop for a suppression motion:

• list_cases — Claude asks what's in your workspace and gets your case roster. You point it at State v. Doe.
• list_evidence — it pulls the inventory: bodycam files, the 911 call, the police report PDF, jail calls.
• search_evidence — you ask what the report claims about the stop. Full-text search across the OCR'd report and diarized transcripts returns the relevant passages.
• read_window — Claude reads the actual transcript window around the officer's stated reason for the stop, with speaker labels and word-level timestamps.
• get_citation — it returns a citation that seeks the exact second of the bodycam video, so what's on your screen matches what Claude quoted.

Notice what's missing: Claude never writes to the case, never drafts a filing, never concludes anything about credibility. Every answer stays tethered to a source moment you can click and verify yourself. That mirrors how the legal-MCP wave is already framing things — CourtListener's integration states plainly that "Claude is not a lawyer" and human judgment remains essential, per LawNext's coverage of the legal MCP launches. Evidence Copilot's own posture is the same: AI output is marked AI-generated and pending human review, and the originals stay immutable.

The practical payoff for a defender is speed with accountability. You can ask Claude to compare the report's stop narrative against the bodycam transcript from your desk, get both sides quoted with timestamps, and then decide — as the attorney — whether it supports a suppression theory. The AI does the retrieval and the citation-finding; you do the lawyering. That division of labor is the whole point of a read-only MCP server, and it's why the five tools are exactly five tools.

Why Read-Only, Org-Scoped, Revocable, Audited — The Security Case

If you connect Claude to your case files, the first question your IT person — or your bar's ethics guidance — will ask is: what stops this thing from doing damage? That skepticism is warranted. Security is by far the most common roadblock in building MCP servers, with 50% of builders citing security and access-control complexity, and 38% saying security concerns prevent them from using MCP more.

The broader MCP ecosystem doesn't inspire confidence on its own. A Fudan University study identified 7,973 live remote MCP servers and found that 40.55% of them exposed tools with no authentication at all. A separate survey found 24% of MCP servers use no authentication whatsoever. For attorneys handling bodycam footage, jail calls, and police reports, that's not an acceptable baseline.

The honest answer to "how do you secure this?" is that you don't rely on the model to police itself. As security researcher Andre Landgraf puts it, model-level defenses against prompt injection "cannot be the foundation on which you build your security architecture." The fix lives at the infrastructure level: authentication, authorization, and scoping. Published best practice says to enforce scopes at the tool level, audit everything, and plan for revocation from day one.

That's the pattern Evidence Copilot's MCP server follows, and it's worth spelling out what each element means in practice:

• Read-only tools — the server exposes exactly five functions: list_cases, list_evidence, search_evidence, read_window, and get_citation. Claude can query your discovery; it cannot delete, modify, or redact anything. Originals stay immutable.
• Org-scoped access — a token sees only the cases and evidence within your organization, not the whole platform.
• Revocable tokens — if a laptop is lost or a team member leaves, the connection dies with the token, immediately.
• Full audit logging — every query Claude makes is logged, the same way every redaction and export in the workspace is audited.

This design also mirrors where the ecosystem is heading. The most-used MCP server type is read-oriented data access — data sources and knowledge bases at 63% — and legal research providers like Thomson Reuters and CourtListener launched their Claude integrations on the same read-heavy pattern. What they don't cover is your evidence media; no defense-side competitor currently offers an MCP integration at all, so there's no established template to copy — which cuts both ways.

One caveat worth stating plainly: read-only and audited is a strong posture, but no tool eliminates risk. The same principle that governs the workspace applies here — AI output is marked pending human review, and the attorney putting their name on the work remains the final check.

Getting Started: Connecting Claude to Your Case in Practice

The theory is fine, but defenders don't adopt tools on theory — they adopt what works on a Tuesday with 300 pages of discovery and a preliminary hearing Thursday. So here's the practical version.

What you need. Claude Desktop or Claude Code, plus an Evidence Copilot account with your case uploaded — bodycam, interrogation footage, jail calls, police reports, whatever the dump contains. Setup mirrors how the legal-research MCPs connect: add the server through Claude's connectors settings and paste in your token. The connection is read-only and org-scoped, covering five tools: list_cases, list_evidence, search_evidence, read_window, and get_citation. That read-heavy design isn't a limitation — it matches how MCP is actually used. Per Zuplo's State of MCP report, the most common server type is data sources and knowledge bases at 63%, and the top-cited value is better context for AI responses.

Three workflows to try first:

• Report-vs-bodycam suppression prep. Ask Claude to pull every window of the bodycam transcript around the search and compare it against the police report's stated timeline. Claude reads the transcript windows, quotes both sides, and you draft suppression questions from the discrepancies — with citations that seek the exact second of footage.
• Quoted excerpts for a case brief. Have Claude search the interrogation transcript for the statements you care about and pull quoted excerpts with page- and timestamp-level citations, ready to drop into your brief.
• Jail-call vs. interview check. Search the jail-call transcript for a statement, then cross-reference what the client actually said in the interview-room footage — Claude can read both windows side by side and show you where the accounts diverge.

The guardrails matter more than the tricks. Everything Claude returns is a draft for your review — marked AI-generated, pending human review, and the product never makes legal conclusions. That's the same posture CourtListener takes with its own Claude integration: "Claude is not a lawyer... human judgment remains essential." Every citation is clickable into the exact source moment, so verification takes seconds, not hours. And security — the top MCP blocker, cited by 50% of builders — is handled at the infrastructure level: revocable tokens, full audit logging, nothing Claude can write back to your evidence.

If you want to see it on your own case, book a 20-minute walkthrough at evidencecopilot.com/book, or start on the Solo plan — $39/user/month, published pricing, no sales wall.