Summary

If you run gbrain from a project directory that has DATABASE_URL in its .env file (common in Next.js, Hono, Supabase, and most Node/bun web apps), bun automatically loads that .env and DATABASE_URL silently overrides the URL in ~/.gbrain/config.json. gbrain connects to the wrong database with no indication that a project .env is involved.

The error message points the user at the wrong place:

Cannot connect to database: Tenant or user not found.
Fix: Check your connection URL in ~/.gbrain/config.json

The URL in config.json is correct. The .env file is the culprit — but nothing says so.

Reproduction

1. Have a working gbrain install with ~/.gbrain/config.json pointing at your brain (e.g. Supabase project A).
2. Create a web project with a .env containing DATABASE_URL=<some other postgres URL> (e.g. the app's own Supabase project B).
3. cd into that project and run gbrain stats.
4. gbrain connects to project B (or fails with "Tenant or user not found" if project B does not exist), with no warning.

Environment

• OS: Windows 11 (also affects any OS where the project .env has DATABASE_URL)
• Bun: 1.x (bun auto-loads .env from CWD by default)
• gbrain: 0.18.2
• Brain backend: Supabase postgres pooler (port 6543)

Root cause

loadConfig() in src/core/config.ts:

const dbUrl = process.env.GBRAIN_DATABASE_URL || process.env.DATABASE_URL;

DATABASE_URL is used by virtually every web framework that talks to Postgres (Prisma, Drizzle, Supabase client, etc.). When bun loads the project .env, it injects DATABASE_URL into process.env before gbrain reads it, so it wins over config.json.

Why this is hard to debug

• The error says "Check your connection URL in ~/.gbrain/config.json" — config.json is fine.
• No indication that an env var is active or where it came from.
• Took ~2 hours to isolate by adding console.log to loadConfig().

Suggested fix

When DATABASE_URL (not GBRAIN_DATABASE_URL) is the active source, emit a one-time warning:

[gbrain] Warning: DATABASE_URL env var is overriding ~/.gbrain/config.json.
         This may be from a project .env file. Use GBRAIN_DATABASE_URL to
         avoid conflicts with application databases.

Or drop the DATABASE_URL fallback entirely — too collision-prone for a tool running inside arbitrary project directories.

Workaround

Set GBRAIN_DATABASE_URL as a persistent env var (takes precedence over DATABASE_URL).

macOS/Linux:

export GBRAIN_DATABASE_URL="<your gbrain connection string>"

Windows:

[System.Environment]::SetEnvironmentVariable("GBRAIN_DATABASE_URL", "<url>", "User")

Also add to the gbrain MCP server env block in ~/.claude.json so it inherits correctly regardless of CWD.