AI Assistant (MCP)

Easy2257 runs a Model Context Protocol server at https://easy2257.com/api/mcp. It lets a producer's AI assistant read that producer's compliance state and answer in plain language: which scenes still need something, what expires next, whether a production is ready to finalize, whether a certificate someone sent them is genuine. It works with Claude, ChatGPT, Gemini, Copilot, Cursor, and any other client that speaks MCP.

It is read-only by design. Nothing reachable through it can create, change, sign, or delete a record, activate a scene, or spend money, and it never returns performer identity data.

Endpoint and discovery

• Transport: Streamable HTTP, stateless. Clients POST JSON-RPC to the endpoint; there is no session to hold between calls.
• Server card: /.well-known/mcp/server.json
• OAuth discovery: /.well-known/oauth-authorization-server and /.well-known/oauth-protected-resource
• Unauthenticated requests answer 401 with a bearer challenge whose resource_metadata points at the protected-resource document, which is what an MCP client needs to start a sign-in on its own.

Authentication

Two credentials, one authorization story. A pasted key and a sign-in token resolve to the same account, the same tools, and the same scoping; nothing downstream can tell them apart.

MCP keys

• Created by the producer at Account, then MCP Keys. Prefix ez_mcp_. Up to ten active keys per account.
• The secret is shown once at creation and stored only as a hash. Rotate or revoke any key at any time. Every key on the account is revoked automatically when the password is reset.
• Send it as a bearer token:

Authorization: Bearer ez_mcp_...

• Keys are live only. They need an active subscription; every plan includes MCP access.

Sign-in (OAuth 2.1)

For clients that connect by sign-in instead of a pasted key (ChatGPT, Claude on the web and desktop, and others with a connectors screen).

The producer approves the connection on an Easy2257 consent screen that names the client and states what it can and cannot read. Connected clients are listed under Account, then MCP Keys with a one-click disconnect, which revokes the grant immediately.

Refusals

Neither response reveals whether a given key exists. Locked and unentitled accounts deliberately answer 403 rather than 401, so an agent refused over billing does not loop on re-authentication.

Connecting a client

claude mcp add --transport http easy2257 https://easy2257.com/api/mcp \
  --header "Authorization: Bearer ez_mcp_..."

{
  "mcpServers": {
    "easy2257": {
      "type": "http",
      "url": "https://easy2257.com/api/mcp",
      "headers": {
        "Authorization": "Bearer ez_mcp_..."
      }
    }
  }
}

Tools

Sixteen tools, all read-only. Results are always scoped to the authenticated account; a record that belongs to someone else comes back as found: false, the same as one that does not exist.

Productions and scenes

Certificates and archives

Account

Solo creators and public codes

Resources

The server also publishes six easy2257:// resources (data-model, closed-scenes, certificates, provenance, removals, solo). They explain the parts of the data model that surprise people: how a person maps to a record, what can and cannot change after a scene closes, when certificates are delivered, why an imported record never upgrades to a verified one, and what happens to a record after it is removed. An assistant that reads them before explaining an odd-looking result will explain it correctly.

What comes back, and what never does

Every tool returns compliance state: ids, public codes (PROD-, SCENE-, LOG-), statuses, counts, timestamps, SHA-256 hashes, and short labels of the caller's own records (a production title, a key name).

No tool returns performer names, dates of birth, ID numbers, images, file paths, document contents, IP addresses, notification bodies, webhook payloads, or anything about a removal-request claimant. Those fields are not filtered out of responses; they are never selected in the first place, and an automated sweep calls every tool against poisoned rows on every build to keep it that way. This is why the connection is safe to point at a shared assistant: there is no identity data in it to leak.

Limits

• One account per credential. A key or token sees the account that issued it and nothing else.
• Foreign lookups are budgeted. verify_record and verify_certificate on records the caller does not own are rate limited per credential. Once the budget is spent, the tool answers the same way whether or not the code exists. Lookups of your own records are not budgeted.
• No paid or attesting path. Scene activation, checkout, attestations, and signatures happen at easy2257.com, by a person, on purpose. An assistant asked to do any of those should point the producer at the site.
• Every call is logged against the credential that made it, and a producer can see when each key or connection was last used.

Producer-facing setup guide: easy2257.com/ai-assistant.

Listed in the official MCP registry as io.github.Agaveis/easy2257-mcp. The registry entry, the connector's privacy statement, and example prompts are at github.com/Agaveis/easy2257-mcp.