Your schema, in your agent's context
The ERD tool your AI coding agent can read.
Design the database in ERDone. Then let Claude Code, Cursor, Copilot — any MCP-capable agent — read it live: the full schema, dialect-correct DDL, the pending migration, the data-loss warnings. Without you ever pasting a schema again.
Set up in about a minute
Create a token in Settings → MCP tokens, then paste one command:
claude mcp add --transport http erdone https://mcp.erdone.net \
-H "Authorization: Bearer <token>"Using Cursor, Claude Desktop, or another client? The same panel gives you a ready-to-paste JSON config. Tokens are named, revocable any time, and shown exactly once. Full details in the docs.
A schema oracle, not a code generator
Your agent already has your codebase in context and writes excellent code. What it doesn't have is authoritative knowledge of your database design — so it guesses at column types, invents foreign keys, and drifts from the real schema. The ERDone MCP server supplies exactly the missing piece: structured schema context, plus the deterministic outputs ERDone already computes correctly — DDL in 8 dialects, baseline-diff migrations, destructive-change checks, schema lint. Byte-identical to what the app generates, because it's the same engine.
Beyond the file in your repo
Every ERDone save already commits an agent-readable SCHEMA.md to your repo. The MCP server covers everything that file can't reach:
Separate repos. Keeping database designs in a different repo from application code is the norm for IP-conscious teams. No file bridges that gap — a network call does.
No local checkout. Cloud and cross-device agent sessions can't read a file they don't have. They can call an endpoint.
Computed on demand. DDL re-targeted to any of 8 dialects, a fresh migration diff against your deployed baseline, lint — computed live at request time, not frozen at your last save.
Fourteen tools, two kinds of answers
Know the design
-
list_diagrams— every diagram the token can read -
get_schema— full schema of one diagram -
list_tables— compact table inventory -
get_table— one table in full detail -
get_related_tables— the FK-neighborhood of a table — context-window-efficient -
search_schema— substring search over names
Compute the outputs
-
get_ddl— any of 8 dialects, identical to the app's Generate SQL -
get_relationships— FK edges with cardinality and actions -
get_views— SQL view definitions -
generate_migration— deployed baseline → current design, data-loss warnings inline -
detect_destructive_changes— just the data-loss check -
lint— ERDone's schema health checks -
list_migrations— browse committed migration history -
get_migration— read one migration file
PostgreSQL, MySQL, MariaDB, SQL Server, SQLite, Oracle, Snowflake, Redshift.
What you'll actually type
The pattern: your agent writes application code against the real schema without you pasting anything — then checks the pending migration for data loss before you deploy.
Diagram → code. Never code → diagram.
The server is read-only by design — there is no write path, and no path by which your application code flows back to ERDone. Every call reads your design live from your own GitHub repo, at the current HEAD, using your own credentials — nothing cached on our servers, nothing stale. Tokens are 256-bit, stored only as a hash, shown once, and revocable instantly. Tier and workspace checks run server-side on every call.
The MCP server is AI reading your design. Want AI reviewing it too? That's Pair with Claude.
Honest note
It's read-only — your agent can read designs, not edit them. It needs a GitHub-connected ERDone workspace, and auth is a pasted token today (a connect button is planned). No ORM model generation yet. Included with Pro ($7/mo) and Team ($12/mo), unlimited calls, no metering; free for 14 days on trial.