Limitations at a glance

This is a summary. Limitations and known issues holds the full list, with workarounds.

• Windows only, with EA on the same machine. A Claude Cowork cloud session has no local MCP server, so it cannot reach EA.
• The server needs EA running with a model open. The server does not start EA by itself. launch_ea exists, and starting an application on someone's desktop is your decision. Where computer use is available the assistant can start EA, open the model, and read and dismiss EA's blocking dialogs, which the server cannot see. Without computer use, you do those steps. For a blocking dialog, you paste a screenshot of it into the conversation so the assistant can read it, then dismiss it yourself.
• Which capabilities you have depends on the client. Computer use, local code execution and browser control are not available everywhere. Without computer use the assistant hands you EA's interface-only steps. Without local code execution the code engines are not available, or the AI computes the result itself at higher token cost and less consistency, depending on the skill. The reporting build, its checks and its refresh run in the server and still work from an existing reporting profile, but the inclusion choice cannot be computed and no Power BI project can be generated. See Capabilities and surfaces.
• The code engines need Python 3 on the machine where EA runs, with pyyaml for the composition engine and pyarrow for Parquet files, where the engine in question uses them.
• One EA instance per seat. Never start a second instance against the same file.
• Plugin updates apply only after a reload. In Claude Code, automatic updates for the plugin's marketplace are off by default: you run the update commands, then reload the plugins or start a fresh session. A plugin uploaded as a file does not update itself; you upload a newer file and restart Claude. See Updating and uninstalling.
• Deletes through EA's automation interface cannot be undone. Take a baseline first.
• A write that EA rejects can be reported as a success at some call sites. Verify by re-reading or with verify_diagram.
• Identifier schemes differ between operations. For example update_attribute takes a GUID, while delete_attribute takes the element and attribute ids.
• ea_validate has one operation, audit.
• The modeling-language listing is not exhaustive, and set_active_mdg records a preference without enforcing anything.
• add_image needs a local .qea or .qeax file. On a server connection or a cloud connection, where computer use is available the assistant can load the image through EA's Image Manager instead.
• Reporting: live or a snapshot depends on how the business layer is delivered. Views in the EA repository's database are live, and Power BI reads them in DirectQuery or Import. A separate reporting database and Parquet files are snapshots, rebuilt in full on every refresh with no history, and Power BI imports them. A scheduled AI agent starts a background refresh, with EA running and the repository open on that machine. Views in the EA database need SQL Server and direct access to it, so they are not available on a local file or, normally, on a cloud connection (Supported repositories).
• Reporting exclusion is by whole package, with no masking inside a kept package. A comma-joined text tag is one value, not a list. A Power BI project built on Parquet files stores an absolute folder path, so moving the folder or the project means generating it again.
• Prolaborate skills need browser control and cannot promise a complete impact analysis.