Erd Designer

ERD Designer CLI

Edit .erd files directly with the bundled CLI. No MCP server, VSCode or running app is required. Requires Node.js 22+.

The CLI ships with this skill at scripts/erd-agent.cjs, relative to the directory containing this SKILL.md. This skill is self-contained (Agent Skills format), so it works from any agent that supports skills — e.g. Claude Code and GitHub Copilot CLI. In the examples below, $CLI stands for the resolved path:

CLI=<directory of this SKILL.md>/scripts/erd-agent.cjs

Workflow

1. Discover tools (names and one-line summaries):

   node $CLI list-tools
   

2. Inspect a tool before first use (full description + JSON schema of its arguments):

   node $CLI describe <tool-name>
   

3. Run a tool against a .erd file:

   node $CLI run <tool-name> --file <path/to/file.erd> --args '<json>'
   

   The result is printed as JSON on stdout. Mutating tools save the file in place.
4. Validate a .erd file (e.g. after external changes):

   node $CLI validate --file <path/to/file.erd>
   

5. Check the design against reality (optional, no MCP tools involved — see below):

   node $CLI erd-diff --file <path/to/file.erd> --from <path/to/other-revision.erd>
   node $CLI db-diff --file <path/to/file.erd>   # needs ERD_DB_URL or --dsn, and pg/mysql2 installed
   

Rules

• Do NOT edit .erd JSON by hand; always go through the CLI so referential integrity (column/share/relation IDs) and the canvas layout stay consistent.
• Omit documentId and filePath in --args; the CLI injects them from --file automatically.
• Only load the schemas you need via describe; do not dump all schemas at once.
• Files are saved as 4-space-indented JSON, the same format the ERD Designer app writes.
• The VSCode extension reflects CLI edits automatically while the file is open; no reload is needed. The browser app (IndexedDB) does not — re-import the file after CLI edits. The Google Drive app reflects CLI edits automatically too, but only when the user has enabled its "Sync Google Drive" toggle (see below).
• Create a brand-new .erd file with run create-document, pointing --file at a path that does not exist yet. databaseType is required and cannot be changed afterwards, so confirm the target database with the user first. The tool never overwrites an existing file.

Typical recipes

• Start a new diagram: run create-document --file <path/to/new.erd> --args '{"databaseType":"postgres"}' (postgres, mysql, mariadb, ms_sqlserver, sqlite, bigquery or snowflake). documentName defaults to the file name without .erd. The response carries the documentId for the tools that follow.
• Add a table with columns: run add-table, then run add-columns-to-table with the returned tableId.
• Relate two tables: look up ids via run list-tables, then run create-relation.
• Review a design: run list-tables (summaries) or run find-table (full detail).
• Generate SQL: run export-ddl.
• Review what changed since a base revision: erd-diff --file <current.erd> --from <base.erd> --format markdown (see below; not an run <tool-name> call, and not a source of truth for documentId — use list-documents / find-document-by-filepath for that).
• Check drift against a live database: db-diff --file <file.erd> (ERD_DB_URL or --dsn required; pg/mysql2 must be installed in the working directory — not bundled).

Schema verification (erd-diff / db-diff / migrate-ddl)

These are separate top-level commands, not agent tools — they don't go through run, don't take --args, and don't touch documentId. They compare .erd schemas against another revision or a live database rather than editing a diagram.

• node $CLI erd-diff --file <path.erd> --from <path.erd> [--format text|json|markdown] — schema-level diff between two .erd revisions. No database connection needed.
• node $CLI db-diff --file <path.erd> — checks the design against a live PostgreSQL/MySQL/MariaDB database. Read-only. Requires ERD_DB_URL (preferred) or --dsn <url>, and the pg or mysql2 driver installed in the current working directory (not bundled with this skill).
• node $CLI migrate-ddl --file <path.erd> [--from <path.erd> | --dsn <url>] [--out <path.sql>] — drafts ALTER statements to close the gap. Never applies them; review the output before running it.

Use erd-diff/db-diff when the user asks "what changed" or "is this in sync with the database" — don't try to answer that by reading the raw JSON.

.erd format (background only)

A .erd file is a single JSON document containing tableViewModels (tables + canvas positions), columnModels / columnShareModels (columns; shares hold the type definition and can be reused by multiple columns), relationViewModels, memos and database settings. IDs cross-reference between these arrays, which is why manual edits are unsafe.

Google Drive / browser users

• Files stored in Google Drive can be edited locally through Google Drive for Desktop: point --file at the synced local path. Tell the user to enable "Sync Google Drive" in the gear menu: with it on, they can keep the ERD Designer tab open while you edit, and your changes appear on the canvas within about 10 seconds. If the toggle is left off, fall back to the old workflow — tell the user to close the tab before you edit, and to reopen the file afterwards.
• The browser app (IndexedDB) is not reachable from the CLI; export the .erd file first, edit it, then import it back.