Evanesca for your AI agents, over MCP.

The endpoint

https://evanesca.app/api/mcp

It speaks JSON-RPC 2.0 over MCP's Streamable HTTP transport, in its stateless form: one JSON-RPC message per POST request, answered with plain JSON. A notification gets an HTTP 202 with no body; there is no session and no server-sent stream (GET answers 405). The methods are initialize, tools/list, tools/call and ping.

Protocol errors come back as JSON-RPC errors. A tool that fails — operator down, credit at zero, a model Evanesca does not serve (the API's model_not_found) — returns a result flagged isError, which the agent can read before deciding what to do.

curl https://evanesca.app/api/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/list"}'

Nine tools

Three read the catalogue and need no key. The other six need a key — or, through a connector, the right shown: chat_completion and create_embedding are billed per token, like the REST API; web_search and web_fetch are billed per call; check_balance and get_profile read your account.

Web search for your agent

web_search and web_fetch call Linkup, a French company — Linkup Technologies SAS, registered office in Fresnes, France — on Evanesca's own Linkup account: the same engine as the chat's web search. Linkup receives only what the tool needs — the query your agent writes, or the address of the page to read — never the conversation, and Evanesca keeps no copy of the queries or the results. Evanesca has not verified where Linkup's servers run; its own privacy policy applies, read on 13 Oct 2026.

• web_search returns up to 8 results, each with the text of its page — not just a snippet — cut at 15,000 characters, so your agent can often answer from the first call. Its optional depth is “fast” or “standard”, the default and more thorough.
• web_fetch reads one page — an http or https address — and returns its text as markdown, cut at 15,000 characters: the follow-up when a result's text is not enough.
• Each call is billed, VAT included: €0.0076 a search, €0.0015 a page — about €7.56 for 1,000 searches. Over MCP that is all you pay for them: your client's own model reads the results. Both need a key — or a connector's web right — and credit above zero.

Connect a client with a key

You need a client that accepts a remote MCP server over Streamable HTTP and lets you add a header. Point it at the endpoint and send your key as Authorization: Bearer. Each client has its own format; these follow the documentation of Claude Code, Cursor and Cline, read on 6 Oct 2026.

claude mcp add --transport http evanesca https://evanesca.app/api/mcp \
  --header "Authorization: Bearer $EVANESCA_API_KEY"

{
  "mcpServers": {
    "evanesca": {
      "type": "http",
      "url": "https://evanesca.app/api/mcp",
      "headers": {
        "Authorization": "Bearer ${EVANESCA_API_KEY}"
      }
    }
  }
}

{
  "mcpServers": {
    "evanesca": {
      "url": "https://evanesca.app/api/mcp",
      "headers": {
        "Authorization": "Bearer ${env:EVANESCA_API_KEY}"
      }
    }
  }
}

{
  "mcpServers": {
    "evanesca": {
      "type": "streamableHttp",
      "url": "https://evanesca.app/api/mcp",
      "headers": {
        "Authorization": "Bearer <EVANESCA_API_KEY>"
      }
    }
  }
}

Claude Code and Cursor read the key from your environment (${EVANESCA_API_KEY}, ${env:EVANESCA_API_KEY}). In Cline's file, paste the key itself in place of <EVANESCA_API_KEY>, and keep that file out of any repository.

ChatGPT and Claude: the OAuth 2.1 connector

ChatGPT and Claude do not take a key in a header: they connect with OAuth 2.1. You sign in to your Evanesca account and approve what the connector may do; no key is copied anywhere. What these clients expect is described in Claude's guide to custom connectors and OpenAI's MCP authentication guide, read on 13 Oct 2026.

1. 01
   Add the connector

   In ChatGPT or Claude, add a custom connector with the endpoint URL, https://evanesca.app/api/mcp. The client reads where to sign you in from the endpoint itself (OAuth protected-resource metadata, RFC 9728).

2. 02
   Sign in and approve

   Evanesca's consent screen names the client, the rights it asks for and where you will be sent back. A client that publishes its identity at an https address — ChatGPT does, and Claude offers it — shows with that address; one that registered itself is flagged as unverified.

3. 03
   Use it, cut it off at any time

   The connector's calls are billed to your credit like any API call. The connection appears among your API keys as “<client> (OAuth)”: revoke it there to cut it off.

Reading the catalogue never needs a right. ChatGPT uses get_profile to recognise the connected account: it returns an ID that changes neither when tokens are renewed nor when you reconnect.

An access token lasts 1 hour; the client renews it with a refresh token that lasts 60 days and is replaced at each use. An old refresh token presented again — after a 30-second grace period — cuts the whole connection, as OAuth 2.1 recommends against theft. Tokens are opaque, Evanesca keeps only their hash, and they work only on the MCP endpoint, not on the REST API. An expired or revoked token gets HTTP 401 with a WWW-Authenticate header, so the client renews it or asks you to sign in again; a tool called without the right it needs returns an error result that names that right.

For client developers: PKCE with S256, the resource parameter (RFC 8707) and the iss parameter on every redirect (RFC 9207). Discovery at https://evanesca.app/.well-known/oauth-protected-resource and https://evanesca.app/.well-known/oauth-authorization-server. A client identifies itself with a metadata document published at its https client_id, or registers itself (dynamic registration, RFC 7591, public clients only). No OpenID Connect.

Keys, connections and billing

The key goes in the Authorization header of every request, as with the REST API. Without it, the catalogue tools still answer — an agent can look before it asks for anything; the others return an isError result asking for a key. An API key carries every right; a connector carries only the ones you approved.

chat_completion reaches the same models as the REST API, all run in Europe, and its result names the operator that answered and the place (Where calls run).

Billing is the REST API's: same prices per token, same credit, same history in your account. A call that fails at the operator is not charged. web_search and web_fetch are billed per call — €0.0076 a search, €0.0015 a page. Once your credit is at zero, the paid tools refuse new calls; calls already running can take it slightly below zero, and the difference comes off your next credit purchase.

Frequently asked questions