Capabilities and surfaces

AI Power Tools does not work through a single channel. The assistant uses four capabilities, and which of them you have depends on the AI client you work in, which this section calls the surface. The section names the four, says what each enables and what happens without it, gives the matrix of surfaces, and shows how to check what your own session has. Elsewhere this document writes "where computer use is available" or "where local code execution is available". This section is the definition.

The four capabilities

Capability	What it gives the assistant	What depends on it
MCP server	The 149 EA operations	All EA work, including building, checking, replaying and refreshing the reporting business layer
Computer use	Seeing and operating desktop applications: EA's own window, and Power BI Desktop	Starting EA and opening a model. Reading and dismissing EA's blocking error dialogs, which EA uses instead of returning errors. Looking at the real EA window to check diagrams, toolboxes and the Quick Linker. EA tasks that exist only in its interface: Save Package as MDG Technology, the Manage Technologies dialog, document templates, and the procedures `ea-help` carries out. Opening and refreshing a generated Power BI project
Local code execution	Running the code engines bundled with the skills, on the machine where EA runs	The diagram advisor and the composition engine; for reporting, the census behind the inclusion choice, the governance gate, writing Parquet files and generating the Power BI project (Code engines); and `ea-com` scripts. The reporting build itself runs in the server. An engine gives the same result every time and uses far fewer tokens than the AI working the same thing out
Browser control	Operating web applications, through a built-in browser or Chrome	All Prolaborate skills. Computer use can see a browser but cannot click in one, so browser control is a separate capability

What each capability enables, and what happens without it

Capability	What it enables	Without it
MCP server	Reading and changing the model, diagrams, analysis, validation, technologies and baselines, and building the reporting business layer, through the 149 operations in MCP server: tools and operations	Nothing that reads or changes the model is available. The server is required for all EA work
Computer use	The assistant starts EA, opens a model, reads and dismisses dialogs the server cannot see, looks at the real EA window to check its own work, and carries out tasks that exist only in EA's interface	The assistant hands you the interface steps and asks what you see. You start EA and open the model, paste a screenshot of any blocking dialog into the conversation so the assistant can analyze it and then dismiss the dialog, look at the diagram, and perform the interface-only tasks. Work that needs only the server continues
Local code execution	The code engines run on the machine where EA runs. Their results are computed, so they repeat exactly, and they cost far fewer tokens	Where a skill has no fallback, the engine's work is not available. Where it has one, the AI does the computation itself, which costs more tokens and is less consistent. Each skill's "Capabilities it uses" table says which applies. For reporting, the server still builds, checks, replays and refreshes the business layer from an existing reporting profile, but the inclusion choice cannot be computed, and no Parquet files or Power BI project can be produced
Browser control	The assistant operates Prolaborate through a built-in browser or Chrome, in a session you have signed in to	The Prolaborate skills are not available. EA work is unaffected

Several statements in this document describe the server, not AI Power Tools as a whole: for example, that the server does not start EA, and that it cannot see an EA dialog. They are true of the server. With computer use the assistant can do both of those things.

Surfaces and capabilities

Surface	MCP server	Computer use	Local code execution	Browser control
Claude Code (desktop app)	Yes	Yes	Yes	Yes (built-in browser or Chrome)
Claude Cowork	Yes (local sessions only, through the Claude Desktop extension; not in cloud sessions)	Yes	No (local sessions run code in an isolated virtual machine, not on the machine where EA runs)	Yes (built-in browser or Chrome)
Claude Desktop chat	Yes	No	No	Yes (built-in browser or Chrome)
VS Code with GitHub Copilot	Yes	No	Yes	Yes (VS Code's integrated browser tools)

Verified against VS Code 1.141.2 and Anthropic's Claude Cowork architecture article, October 2026. Surfaces gain capabilities over time.

Claude Cowork. Claude Cowork runs in the cloud by default. A cloud session has no local MCP server, so it cannot reach EA. A local session reaches the server through the Claude Desktop extension. Its code runs in an isolated virtual machine, not on the machine where EA runs, so an engine that needs EA's automation interface or Windows cannot run there. An engine that needs only Python can use only the inputs and outputs in a folder you have connected on the desktop.

VS Code with GitHub Copilot. Browser control comes from VS Code's integrated browser tools, which are on by default. A page the assistant opens runs in an isolated session. To work in a page you have signed in to, open it in the integrated browser and share it with the assistant.

Tool limit in VS Code. VS Code allows at most 128 tools in one chat request. The MCP server uses 7. The limit matters only when many other tool servers are enabled alongside it. See VS Code extension.

Claude Cowork and cloud sessions: reading and reporting

In Claude Cowork and in other cloud sessions, AI Power Tools is a reading and reporting surface, not a modeling surface. The assistant works with what the server or the code engines have already written to a folder connected on the desktop: exports, Parquet files, generated Power BI projects and reports. For example, from the Parquet files of a reporting database it can produce a portfolio roll-up without EA running. To model, analyze or validate in EA, use a surface that has the MCP server in the session: Claude Code, a local Claude Cowork session, Claude Desktop chat or VS Code. For what leaves your machine in a cloud session, see Security and data handling.

Where the MCP server comes from. On Claude Desktop chat, Claude Cowork (in local sessions) and the Code tab of the Claude desktop app, the server is the Claude Desktop extension (Claude Desktop extension). The skills come from the Claude plugin installed on your claude.ai account (Claude plugin), which carries no server. Claude Code in a terminal does not load Claude Desktop extensions, so it has no server from these packages. In VS Code the server comes from the VS Code extension (VS Code extension).

Surfaces gain capabilities over time, so this document names a surface only in this matrix and in examples, and otherwise writes in terms of capabilities. "Claude Cowork" is the Claude product. Microsoft 365 Cowork is a different product and is not covered here.

Runtime requirements for the code engines. Local code execution runs Python on the machine where EA runs:

Python 3.
pyyaml, for the composition engine's notation bindings.
pyarrow, only for delivering the reporting business layer as Parquet files. It is deliberately not bundled, so install it when you choose Parquet files.
pywin32, only for scripts the AI writes to drive EA's automation interface directly, in ea-com and in ea-mdg-deploy's automation checks. No reporting step needs it.
Everything else is Python's standard library.

The AI decides how to use what the surface offers

AI Power Tools does not detect your surface and switch to a fixed mode. The assistant sees which capabilities its session has and decides how to use them:

With computer use it may start EA, read a dialog, or look at a diagram in EA's window. Without it, it asks you to do or describe those things.
Where local code execution is available, a skill directs the assistant to run an engine. The engine's result is computed and repeatable, but the AI chooses when to run it and what to do with the result.
Where browser control is available, the assistant operates the web application.

The same request on two surfaces can therefore take different paths to the same result. In the labels of Where the AI decides and where AI Power Tools decides, using a capability is an AI decision, and what the server or an engine computes is AI Power Tools'. Skills mark the two cases that depend on capabilities: AI with computer use; otherwise you and Engine. Check results as each skill's accuracy checks describe, whichever path was taken.

Models and GitHub Copilot

The model guidance in this document is written for Claude models. VS Code with GitHub Copilot can also use AI models that run on the device. Skills whose work is mostly done by code engines depend least on the model. Skills that rely on AI judgment depend most. This document makes no rating claims for on-device models.

How to tell which capabilities your session has

Ask the assistant. It knows which tools its client has given it. A prompt such as "Which of computer use, local code execution and browser control do you have in this session, and is the EA server connected?" is enough. To confirm rather than take its word:

MCP server. Ask it to call ping, then get_repository_info, as in Verifying an installation.
Computer use. Ask it to describe what is on the EA window. If it can, it has computer use. If it asks you to describe the screen instead, it does not. The client typically asks you to approve each application before the assistant operates it.
Local code execution. Ask it to report the Python version on the machine and whether the packages in the requirements list above are installed.
Browser control. Ask it to open a web page and read its title.

A capability the matrix shows as available can still be switched off or limited in your client's settings or by your organization's policy, so test rather than assume. Each skill's "Capabilities it uses" table in Skills says what the skill needs and what changes without each capability.