← Back to blog

Eurobase Now Speaks MCP: AI-Native Backend Operations, Sovereign by Default

Stefan Gimeson··5 min read
Eurobase Now Speaks MCP: AI-Native Backend Operations, Sovereign by Default

AI coding assistants have spent the last two years getting better at writing code. The next leap is them getting better at operating the systems that code runs on. As of this week, Eurobase ships a hosted Model Context Protocol (MCP) server so your assistant can do exactly that — and it does it without leaving European jurisdiction.

What Just Shipped

A live MCP endpoint at https://mcp.eurobase.app/mcp, plus a "Connect" tab in the console that generates the right snippet for every major AI IDE: Claude Code, Codex, Cursor, Windsurf, or any client that speaks the open Streamable-HTTP MCP transport.

Once configured, your assistant gains 15 tools that map onto the platform API:

  • Projects — list and inspect every Eurobase project you can access
  • Database — list tables, describe schemas, run SQL, create tables
  • Auth — list end-users registered in a project
  • Storage — list files, generate signed download URLs
  • Vault — list, read, and write encrypted secrets
  • Functions — list and invoke edge functions
  • Status — health-check the API

These aren't suggestions the model writes for you. They're tool calls the assistant executes in real time. Asking "how many users signed up to LexVault this week?" no longer means generating a SELECT for you to copy into the SQL editor — the assistant runs the query itself, gets the number back, and tells you the answer.

Why This Matters

There's a real difference between teaching an AI about your project and letting it operate on your project. A CLAUDE.md or .cursorrules file (which Eurobase also generates for you, schema-aware) gives the model context. MCP gives it agency.

That distinction shows up the moment you're doing real work. The assistant can:

  • Inspect schema before writing migration code, instead of guessing column names
  • Verify a query returned what it expected, not just that it compiled
  • Rotate a Vault secret as part of fixing a leaked credential, in the same session
  • Run a function and read its output back into the conversation
  • Cross-check user counts and storage usage before recommending a plan upgrade

The agent stops being a code generator and starts being a junior engineer with read access to your live system.

The Sovereignty Angle

Most MCP servers shipped this year run on US infrastructure. The leading vibe-coding platforms route their tool calls through AWS, Cloudflare, or Vercel. Even when the model itself is hosted elsewhere, the tool plane is American — and every list-tables call, every SQL statement, every secret read is a request that could be subpoenaed under the CLOUD Act.

Eurobase's MCP server runs on Scaleway in Paris, behind the same EU-only ingress as the rest of the platform. Tool calls never traverse a US-controlled network path. The only data your AI vendor ever sees is what the model itself reads in the conversation transcript — not the underlying queries, schemas, or secrets.

If you've moved your backend to Eurobase for sovereignty reasons, your AI tooling can now match. No bolted-on US dependency at the agent layer.

Personal Access Tokens, Designed for Bots

We didn't want to ship MCP and tell users "just paste your console JWT into a config file." JWTs are a bad fit for tooling: short-lived, no naming, no revocation, and they carry every claim your account has — including superadmin if you happen to be one.

So we shipped Personal Access Tokens alongside the MCP server. Mint one in Account → Personal Access Tokens with a name ("my laptop", "ci-prod") and an optional expiry. The plaintext is shown once on creation; only the SHA-256 hash is persisted.

PATs are deliberately scoped down from a console login:

  • Authenticate as you across every project you own or are a member of — full SDK and platform-API surface, scoped by RLS and project roles like any other request.
  • Never carry superadmin rights, even when minted by a superadmin. Platform admin endpoints are unreachable through a PAT — a leaked token can't reach the cross-tenant surface.
  • Cannot mint other tokens. Token creation requires an authenticated console session. Limits the blast radius of a single leaked PAT.
  • Cannot change passwords or delete the account.

Revocation is immediate and per-token. Last-used timestamps surface in the UI so inactive tokens are easy to spot and clean up.

The MCP-vs-SDK-vs-Migrations Rule

Once an agent has both MCP tools and SDK awareness, the question becomes when to use which. The generated CLAUDE.md and AGENTS.md files now embed an explicit rule the assistant reads on every session:

  • SDK — code written into the application that runs in production at request-time, scoped by end-user RLS. Use it for everything the deployed app does on behalf of its users.
  • MCP — tool calls the agent makes during a coding session, scoped by your PAT. Default to read-only: SELECT, describe, list. Use it to inspect the live system.
  • Migrations — durable schema changes, version-controlled, reviewed in PR, replayed in CI on every environment.

Concretely: a schema change of any kind goes in a migration file under migrations/, never as a silent mcp__eurobase__db_execute_sql call against the live DB. If the change isn't checked into git, it doesn't exist on staging or in any teammate's branch — and that path produces real outages.

The rule is tight enough that the assistant follows it without prompting. We tested it.

Try It in Two Minutes

1. Sign in to console.eurobase.app.

2. Open Account → Personal Access Tokens and click New token.

3. Open any of your projects, go to Connect, and copy the snippet for your IDE.

4. export EUROBASE_PAT=eb_pat_... in your shell, then run the snippet.

That's it. Open Claude Code (or Cursor, or Codex, or Windsurf) and ask it about your project. It will reach in, look around, and answer.

What's Next

Tool coverage is where we expect the most rapid iteration based on beta feedback. Webhook management, edge-function deployment, and signed-URL upload are queued. Read-only audit-log access is too — useful for "what changed in the last hour?" debugging conversations.

If you're a beta tester building with AI, this is the piece you've been asking for. If you've been waiting for sovereign AI tooling that doesn't quietly route through the US, it's here.

Sign in to your console to mint a PAT and connect, or sign up free if you don't have a project yet.