Why MCP is the right shape for an agent backend

The Model Context Protocol gives every agent the same connection contract — /.well-known/oauth-protected-resource, audience-bound JWTs, scope-gated tools. Build once, work everywhere.

For an agent backend, that means we can ship a single MCP endpoint and Claude Code, Claude Desktop, Cursor, and VS Code all talk to it without bespoke integrations.

The tool surface stays tight

connect0 ships five MCP tools: info, account.create, project.create, billing.start_checkout, sh. That's the whole API. Everything an agent does — pick an account, run a build, open a billing portal — is a composition of those primitives.

Compare that to the "wrap our REST API in MCP one-to-one" pattern: fifty tools, none of them well-described, the agent's tool selector swamped. We followed Cloudflare's enterprise MCP guidance — fewer, well-described tools consistently outperform exhaustive API mirrors.

What the backend actually does

Behind those five tools, connect0 is doing a lot:

• OAuth 2.1 with PKCE, audience-bound JWTs signed with rotating ES256 keys.
• Cloudflare Sandbox containers, one per project, keyed by the project's UUID.
• R2-backed /workspace so state survives restarts.
• Stripe Checkout + Billing Portal for subscriptions.
• Server-side audit log of every sh invocation.

You don't see any of that. You see sh "pnpm test" and the result.