Scribase MCP¶

A hosted Postgres backend an AI agent cannot wreck. Scribase gives Claude Code, Cursor and any MCP client 44 tools over your projects: reads run immediately, every write waits for an explicit confirm, and schema changes must pass a row-level-security isolation test before they can ship.

Server type: Remote (stdio or Streamable HTTP at /mcp)
Auth: personal access token or org API key, audited per call
Endpoint: https://api.scribase.com/mcp
Tools: 44 (schema, data, previews, backups, policies)
Pricing: vendor pricing (scribase.com)
Category: DevOps / Database
Built by: Scribase (scribase.com)

Why This Matters for Operators¶

The moment you let a coding agent touch a production database you inherit a new failure class: plausible-looking schema changes that break isolation between tenants or destroy data. Scribase addresses the specific risk with two mechanisms, confirm-gated writes that return the exact request before execution, and an RLS proof system where schema.apply only accepts a token from policy.test, which proves the isolation matrix and then deliberately breaks each guarantee to show the proof works.

Preview branches add the last missing layer: changes land on a time-limited sanitized copy of production, and merges are foreign-key safe and preceded by a backup.

Tools & Capabilities¶

Tool area	Purpose
Reads and writes	44 tools over projects; writes confirm-gated
Schema management	schema.apply requires a passing policy.test token
RLS policy testing	policy.test proves the isolation matrix before ship
Preview branches	TTL-swept sanitized copies for safe iteration
Backups and merges	Foreign-key safe merges, backups before every merge

Only schema.apply and policy.test are named in the published docs; the remaining tools are served from the endpoint.

Installation¶

claude mcp add scribase --transport http https://api.scribase.com/mcp

An install-once setup plus editor integration flow is documented on the server page.

Configuration¶

{
  "mcpServers": {
    "scribase": {
      "type": "http",
      "url": "https://api.scribase.com/mcp"
    }
  }
}

Business Relevance¶

Founders let coding agents build internal tools against a database that cannot be silently broken
CTOs get schema changes gated behind provable RLS isolation
Data teams preview changes on sanitized branches before they touch production
Agencies ship client data work with an audit trail per API call

Integration with CorpusIQ¶

Scribase is the safe landing zone for data CorpusIQ connectors already pull: QuickBooks financials, Stripe charges and Shopify orders can be materialized into Scribase-managed Postgres for agent-driven analysis, where the confirm gate prevents an over-eager assistant from mutating source data.

The CorpusIQ GitHub connector closes the loop on delivery: schema changes proposed by an agent in Scribase can be reviewed as normal pull requests, with the RLS proof attached as evidence.

Limitations¶

New listing, no track record yet
Only two of 44 tools are named in public docs
No free tier disclosed on the directory listing
Database-only scope, no application hosting
Requires comfort with Postgres concepts to configure well

FAQ¶

How does it stop an agent from breaking production?¶

Writes are confirm-gated, meaning the agent must approve the exact request, and schema changes require a passing RLS isolation test first.

What is the RLS proof?¶

policy.test proves the isolation matrix, then breaks each guarantee on purpose to show the proof can fail. Only that token unlocks schema.apply.

Can I try changes safely?¶

Yes. Preview branches are time-limited sanitized copies of production, and merges are foreign-key safe with a backup first.