Connect an AI assistant

Help  ·  Claude, ChatGPT, Claude Code, Cursor and VS Code

You can connect your own AI assistant to Useful EA. It can then answer questions about your capability models, Benefit Dependency Networks, maturity assessments, application portfolios and architecture repositories, such as "which capabilities are low maturity but high priority?" or "what changes if we fund the new plan?"

The assistant signs in as you. It reads only what you can already see in the tools, through the same access checks. It can't read another person's or company's models. It can't change anything in the Capability Modeler, BDN, EA Maturity Assessment or Application Portfolio. In an Architecture Repository, if you allow it, it can propose changes: nothing changes until someone who can edit the repository accepts them.

On this page

Before you start

The server address is the same for every assistant:

https://mcp.useful-ea.com/mcp

AI apps rename their menus often. If a step below doesn't match what you see, look for "connectors", "apps", "integrations" or "MCP servers" in the app's settings.

Claude (web, desktop and mobile)

  1. In Claude, open Settings → Connectors.
  2. Choose Add custom connector.
  3. Name it Useful EA and paste the server address. Leave the advanced settings empty.
  4. Choose Add, then Connect, and sign in to Useful EA.

Once it's added on the web or desktop, the connector also works in the Claude mobile apps. On a Team or Enterprise plan, an owner may need to add the connector for the organization first; you then connect your own account to it.

ChatGPT

  1. In ChatGPT, open Settings → Apps and, under Advanced settings, turn on Developer mode.
  2. Choose Create app (or Create connector).
  3. Name it Useful EA, paste the server address, and choose OAuth for authentication.
  4. Choose Create and sign in to Useful EA.

Developer mode is available on some ChatGPT plans only, and on business plans an administrator may have to allow it.

Claude Code

Add the server from a terminal:

claude mcp add --transport http useful-ea https://mcp.useful-ea.com/mcp

Then start Claude Code, type /mcp, choose useful-ea and Authenticate. Your browser opens the Useful EA sign-in page. Add --scope user to the command to use the server in every project.

Cursor

Open Cursor Settings → MCP (or Tools & Integrations) and choose Add a custom MCP server, or add this to ~/.cursor/mcp.json yourself: 

{
  "mcpServers": {
    "useful-ea": {
      "url": "https://mcp.useful-ea.com/mcp"
    }
  }
}

Cursor shows a Needs login or Connect button next to the server. Choose it and sign in.

VS Code (GitHub Copilot)

Open the Command Palette, run MCP: Add Server, choose HTTP, paste the server address and name it useful-ea. Or add this to .vscode/mcp.json in your workspace: 

{
  "servers": {
    "useful-ea": {
      "type": "http",
      "url": "https://mcp.useful-ea.com/mcp"
    }
  }
}

VS Code asks to sign in the first time Copilot uses the server, in agent mode. Choose Allow, then sign in to Useful EA.

What the assistant can read

Only what you can open yourself, and only in the tools you have:

ToolWhat it can read
Capability ModelerYour models, their capability trees with maturity, cost and stewards, strategies, view names, and the priorities report.
BDNYour networks: items, connections and views; each program's and project's score, rank and place against the funding line; and a comparison of two weight sets.
EA Maturity AssessmentYour assessments, their scores and action plans, and comparisons between assessments.
Application PortfolioApplications, their custom fields and their links to capabilities.
Architecture RepositoryRepositories you can open, their metamodel, elements and relationships. It can also check proposed changes against the metamodel without saving anything.

It never reads:

It can't create, change or delete anything itself. The only thing it can submit is a proposal to a repository, if you allowed that when connecting. Each connection may make up to 60 requests a minute and 2,000 a day. If you need more, contact sales@useful-ea.com.

Propose changes to a repository

An assistant can read your own material (documents, spreadsheets, intranet pages, diagrams), or a company's public sources, and propose what it finds to an Architecture Repository: new elements, changes to ones you have, relationships, links to records in the other tools, and archiving. Every finding comes with its evidence: the source, where in it, an exact quote of up to 3,000 characters, a one-line reason, and how confident the assistant is.

A proposal changes nothing. It waits on the repository's Proposals tab, where anyone who can edit the repository reviews each finding and accepts or rejects it. Undecided findings expire after 30 days. Accepted elements say which assistant proposed them, for whom, and who accepted them.

What you need

Your AI app may ask you before each call that submits a proposal. Either way, the proposal still waits for a person to review it in the Repository.

Two ready-made prompts

A connection that can propose also offers two prompts: named instructions the assistant follows step by step, so you don't have to write them. Each one reads the repository's metamodel first and uses only its types, looks for what's already there so it proposes an update instead of a duplicate, checks the whole set against the metamodel, and only then submits it.

PromptWhat it doesWhat you give it
discover_into_repository
Discover into a repository		 Reads the material you give it and proposes what it describes. repository (name or id) and focus (what to look for), both required; sources (links or document names; otherwise whatever you attach) and conventions (house rules on naming or scope), optional.
profile_company
Profile a company from public sources		 Builds a profile of a company from its annual report, its own web sites and reliable news. repository and company, required; website, annual_report, focus and as_of (ignore older sources; otherwise the last three years), optional.

Unless you give it a focus, profile_company covers the company itself, its main business units and subsidiaries, its products and services, markets and customer segments, main locations, stated strategy, goals and major programs, the business capabilities those imply, its main partners and suppliers, and the major systems it is publicly known to use. It needs an assistant that can search and read the web.

Where to find the prompts: in Claude Code, type / and choose /mcp__useful-ea__discover_into_repository or /mcp__useful-ea__profile_company. Other apps list a connector's prompts in their prompt, attach or / menu. If your app doesn't show prompts, ask in your own words, as in the examples below, and say "propose it to the repository for review".

Limits

Worked examples

Your own material: an application inventory

You have a spreadsheet of applications and an intranet page describing who owns them, and a repository called Retail estate. Attach the spreadsheet, choose discover_into_repository, and fill in:

Without prompts, ask:

Using Useful EA, read the attached Application Inventory 2026.xlsx and
https://intranet.example.com/it/systems. Find the applications, the business
units that own them, and the capabilities they support, and propose them to my
Retail estate repository for review. Use each application's product name, not
its project code.

The assistant reads the repository's metamodel, searches the repository for each application it finds, checks its findings, and submits them. It then tells you how many it submitted, what it left out and why (for example, a row with no owner when the metamodel needs one), and where to review them. Open the repository's Proposals tab to review.

Public sources: profiling a company

You're preparing for a new client, Northwind Traders, and want a first picture of the business in a repository called Northwind discovery. Choose profile_company and fill in:

Without prompts, ask:

Using Useful EA, build a profile of Northwind Traders (https://www.northwind.example)
from its latest annual report, its own web sites and reliable news from the last
three years, and propose it to my Northwind discovery repository for review.
Quote every source exactly.

The assistant first makes sure it has the right company, and asks you if the name matches several. It prefers the company's own report and web sites, then established news with a named publisher and date, and doesn't use forums, social media or undated pages. A fact the company states itself is high confidence; one reported by a single publication is medium; one it inferred, such as a capability implied by a product line, is low, and says so. It proposes nothing about individual people beyond their published role as an officer of the company.

Check the evidence before you accept. Assistants can misread a source or quote the wrong passage. In the review, open a few sources and check that each quote is really there, especially for low- and medium-confidence findings. Accept by filter takes only findings that have been on your screen.

Keep your data safe

When you connect an assistant, the data it reads goes to that AI company under your agreement with them, not ours. Check that your company allows it.

Text in models can be written by other people. Collaborators, students and clients can type instructions into a model's names and descriptions, aimed at your assistant. Useful EA's connection can only read, but your assistant may have other tools that can browse the web or send email, and we don't control those. Don't let an assistant send data anywhere because text inside a model told it to, and review what it does with other tools.

Disconnect an assistant

In any Useful EA tool, open your profile and go to Connected AI apps. Each assistant you've connected is listed with when it was connected and last used, and whether it reads only or can also propose changes. Choose Disconnect: it loses access straight away. Removing the server inside the AI app alone doesn't end its access, so disconnect it here too.

Stop changes takes away an assistant's permission to propose changes and keeps it connected for reading. It applies from the assistant's next request. To allow changes again, disconnect and connect again with the box ticked. Proposals it already submitted stay in the repository for review.

For company administrators

In the organization portal, open Settings → AI assistants. The setting applies to every member of the company:

Lowering the setting asks you to confirm and takes effect at once: No access ends every member's connection, and Read only takes the permission to propose changes away from every connection that had it. Each connection and disconnection appears in the company's audit log, and accepted proposals show in each repository's history.

If it doesn't work

Still stuck? Email support@useful-ea.com and say which AI app you're using.