SeedBase
Test data your AI agent can actually generate
AI coding agents are great at writing code and bad at inventing realistic relational data. They hallucinate IDs that break foreign keys, or write a seed script that rots on the next migration. Connect SeedBase over MCP and your agent generates real, foreign-key-consistent data from your schema, straight into your database.
Why agents need a real tool for this
- Ask an agent for "100 orders" and it makes up
customer_idvalues that point at customers who do not exist. Load that into a real DB and it stops at the first constraint. - The alternative, having the agent hand-write a seed script, breaks the moment your schema changes.
- SeedBase gives the agent a tool that already knows how to resolve foreign keys, order inserts and keep values coherent. The agent asks; SeedBase generates.
Connect it in one command
SeedBase is agent-native: it exposes an MCP server, so the agent calls it as native tools. You need a SeedBase API key (shaped dr_sk_..., created at seedba.se under Settings, API keys).
Hosted HTTP endpoint (simplest)
claude mcp add-json seedbase '{"type":"http","url":"https://seedba.se/mcp","headers":{"Authorization":"Bearer dr_sk_..."}}'Local stdio via npx (no install)
claude mcp add-json seedbase '{"type":"stdio","command":"npx","args":["-y","-p","@seedbase/client","seedbase-mcp"],"env":{"SEEDBASE_API_KEY":"dr_sk_..."}}'Cursor, Windsurf and any other MCP client take the same HTTP URL or stdio command in their MCP config.
What your agent can do
| list_projects | List your SeedBase projects, so the agent can pick the right schema. |
| get_ddl | Read a project's schema as CREATE TABLE statements, per dialect. |
| generate_test_data | Generate a foreign-key-consistent dataset and return it as SQL, ready to load. |
So a prompt like "fill my dev database with realistic data for the orders schema" becomes: the agent lists projects, reads the DDL, generates a dataset, and hands you loadable SQL. No hallucinated rows, no throwaway script.
Works with your stack
- Claude Code, Cursor, Windsurf, and any MCP client.
- Headless / CI: a Claude skill and a plain REST API cover containers and pipelines where you only have an API key.
- Editors: VS Code and JetBrains plugins generate and push without leaving the IDE.
Let your agent generate real test data.
Create an API key, add the MCP server, and ask your agent to fill a database. FK-consistent, loadable SQL, free tier, no card.
Get an API key, freeMore: MCP & API docs · Django test data · seed a Django DB · home