Short answer: there are three different ways to use Claude for M&A due diligence, and they solve different problems. A skill is a packaged set of instructions that teaches Claude how you want a task done, such as how to read a CIM and cite it. A plugin, like the private equity templates Anthropic published in 2026, bundles skills, connectors and sub-agents into one installable unit. MCP, the Model Context Protocol, is the open standard that lets Claude reach into outside systems, for example a virtual data room, so it can search and read documents where they live. Skills shape how Claude works, plugins package a whole workflow, and MCP decides what Claude can see. None of them verifies a number for you.
This page is the hub for our Claude and diligence articles. It explains the three options, says when each fits, and links to the deeper guides. Product facts were read on Anthropic's own pages on 8 November 2026 and are dated below.
The three ways at a glance
| Skills | Plugins (such as the private equity templates) | MCP connection to a data room | |
|---|---|---|---|
| What it is | A folder of instructions, with optional scripts and templates, that Claude loads when a task matches | A bundle that can include skills, agents, hooks and MCP servers, installed as one unit | A connector that exposes a system's actions (list rooms, search, read a file) to Claude |
| What it changes | How Claude does a task | Which workflow you start from | What Claude can access |
| Who builds it | You, or anyone sharing a skill | Anthropic, a vendor or your firm | The data room vendor, usually |
| Typical effort | An afternoon to write; more to test | Needs setup and adaptation | Needs the vendor to offer it, plus admin approval |
| Main risk | A vague skill produces confident, uncited output | A template is a starting point, not a finished product | Permissions, confidentiality and what leaves the room |
| Detailed guide | Using Claude for diligence | Plugin vs turnkey tool | Connect a data room via MCP |
Way 1: Claude skills for due diligence
Anthropic's documentation, read 8 November 2026, describes Agent Skills as modular capabilities that extend Claude. Each skill packages instructions, metadata and optional resources such as scripts and templates. Claude keeps only each skill's name and description in context at the start, and loads the full instructions when a request matches. In practice, you can install many skills without filling the context window.
Anthropic's help center, read the same day, says skills are available on the Free, Pro, Max, Ultra, Team and Enterprise plans and require code execution to be enabled. Check your own plan and your organisation's admin settings before building a process around them, because those details change.
What a diligence skill is good for:
- A reading discipline. "For every figure, quote the source line and page. If there is no line, list it as unsupported."
- A repeatable format. The same sections, in the same order, on every CIM, so the third deal of the month gets the same read as the first.
- A question list. Turning flagged gaps and contradictions into questions for the seller.
- Domain definitions. Your firm's own meaning of recurring revenue, add-back or customer concentration.
What it cannot do: a skill is text. It cannot make a model read a page it was not given, and it cannot stop a plausible invented figure by itself. The check is still you, spot-checking quotes against documents. Our article on 20 Claude agents for M&A diligence shows what small, single-purpose agents look like when each one is built to cite or cut.
A practical starting skill is short. State the role, the documents it will receive, the output format, and the single rule that matters most: cite or cut. Then test it on a deal you have already closed and compare the output with what you know to be true.
Way 2: the private equity plugin and agent templates
In 2026 Anthropic published a set of agent templates for financial services, including private equity and diligence workflows. Our comparison of the plugin with a turnkey tool summarises Anthropic's announcement. The key point is that these are reference architectures that package skills, connectors and sub-agents, and which you adapt to your own conventions and data. They are not a hosted product you log in to.
Anthropic's plugin documentation, read 8 November 2026, describes a plugin as a directory of skills, agents, hooks, MCP servers or other components that is installed and loaded as one unit, usually from a marketplace.
When a plugin fits:
- Your firm has someone technical who can adapt and maintain it.
- You already have data tools to connect, and the connectors are part of the point.
- You want a shared workflow across a team, rather than each person writing prompts.
When it may not:
- You are a solo searcher or a two-person fund with no one to maintain it.
- You need a finished, cited brief from a CIM today.
- Your firm's rules on third-party software and data would need review first.
A template saves you from starting from nothing. It does not remove the work of testing it on your own documents, and it does not change the verification rule: every figure still needs a source.
Way 3: MCP to a data room
MCP is described on modelcontextprotocol.io, read 8 November 2026, as an open-source standard for connecting AI applications to external systems. For diligence, the interesting case is a data room that publishes an MCP connection. Claude can then call actions such as list rooms, search documents or read a file during a chat, instead of you downloading files and pasting them in.
Our article on connecting a data room to Claude via MCP lists what several data room vendors published as of 9 October 2026. Two points from it matter here.
- The connector lives on the vendor's side. If your data room does not offer one, you cannot add it from outside without exporting documents, and exporting from a confidential room is often exactly what room rules forbid.
- Permissions decide the quality of the setup. A connector that inherits each user's existing access is different from one that holds a single broad token. Test it with two accounts that have different rights and see whether the one with fewer rights can reach what it should not.
MCP changes what Claude can see. It does not change what Claude does with it, which is why it pairs naturally with a skill that enforces citations.
Which should you start with?
| Your situation | A sensible starting point |
|---|---|
| Solo searcher, a few CIMs a month | A short skill, plus strict citation checks by hand |
| Small sponsor with a technical partner | A skill first, then adapt a plugin template once the skill is working |
| Seller or buyer already in a room that offers MCP | Test the connector with two permission levels, then add a skill |
| Eighty-file room and a team that must produce the same read every time | A system built for it: a plugin you maintain, or a turnkey tool |
| No one to maintain anything | A turnkey tool that returns a cited brief |
If you are choosing between Claude and another model for the reading itself, our Claude vs ChatGPT comparison argues that the model matters less than forcing every claim to cite its source.
Confidentiality checks before any of this
Whichever route you take, confirm these in writing before real deal documents go in:
- Your NDA and process letter allow processing the documents with an AI service.
- The plan's data terms. Read what your Claude plan, and any connector vendor, say about storage, retention and use of data for training.
- Who has access. Shared skills and plugins can carry instructions, scripts and connectors. Read what you install before you run it on confidential material.
- A log. Keep a record of what was uploaded, to which tool and when.
Where Deal OS fits
Deal OS is our own product, built on the cite-or-cut rule rather than on a plugin you assemble. You upload a CIM and get a brief in which each figure is traced to its source page. Its pricing page states that the first CIM is free, then $99 per CIM, or plans from $249 per month, in USD excluding VAT. If you have the time and skill to build your own workflow, the three routes above are legitimate and cost less cash but more of your time. If you want the result without the build, a turnkey tool is the shorter path. See getdealos.tech.
Frequently Asked Questions
What is a Claude skill for due diligence? It is a packaged set of instructions, optionally with scripts and templates, that Claude loads when a task matches. For diligence, a skill typically defines how to read a CIM, the output format and the rule that every figure must cite its source. It shapes the work but does not verify it.
What is the Claude private equity plugin? It is one of Anthropic's 2026 agent templates for financial services. A plugin bundles skills, connectors and sub-agents as one installable unit, and you adapt it to your own workflow and data. It is a reference architecture, not a hosted diligence product.
Can Claude connect to my data room? Only if your data room vendor offers an MCP connection and your administrator enables it. The connector sits on the vendor's side. Test it with accounts of different permission levels before relying on it.
Are Claude skills, plugins and MCP the same thing? No. Skills change how Claude performs a task, plugins package a whole workflow including skills and connectors, and MCP controls what outside systems Claude can reach. They can be combined.
Is it safe to use Claude on confidential deal documents? That depends on your NDA, your plan's data terms and your permissions setup. Read all three before uploading, and keep a log of what went where. None of the three routes removes the need to check figures against the documents.
Which is best for a small search fund? A small fund can reasonably start with a short skill and strict citation checks, because it needs no connector and no maintenance team. If building and maintaining that is not realistic, a turnkey tool that returns a cited brief is the simpler option.