API MCP server

What it is

The server is the REST API as tools, one per endpoint. It takes the same keys, charges the same prices and answers the same objects as /v1, so everything an agent does here is a /v1 request: listed in your console's Logs under its key and paid from your API balance. Use it when an agent builds your integration, or renders in a pipeline nobody watches.

Endpoint

Streamable HTTP at https://api.elyum.ai/v1/mcp, stateless, with a bearer token on every request. A client that supports MCP sign-in finds our authorization server on its own: the first answer, a 401, names it. The others send an API key.

The studio's MCP server

Claude Code

One command in your terminal. Claude Code signs in through our consent page (OAuth); or pass an API key in a header.

claude mcp add --transport http elyum-api https://api.elyum.ai/v1/mcp

# or, with an API key:
claude mcp add --transport http elyum-api https://api.elyum.ai/v1/mcp --header "Authorization: Bearer ek_live_…"

Claude.ai and Desktop

Settings → Connectors → Add custom connector. Paste the URL; Claude signs in through our consent page (OAuth), where you pick scopes and a daily cap in dollars.

Name:  Elyum API
URL:   https://api.elyum.ai/v1/mcp

Cursor

Click the button (deep link) or add it to ~/.cursor/mcp.json. Leave out headers and Cursor signs in through our consent page instead.

{
  "mcpServers": {
    "elyum-api": { "url": "https://api.elyum.ai/v1/mcp", "headers": { "Authorization": "Bearer ek_live_…" } }
  }
}

Codex

Codex CLI and IDE read the key from the environment variable named here.

# ~/.codex/config.toml
[mcp_servers.elyum-api]
url = "https://api.elyum.ai/v1/mcp"
bearer_token_env_var = "ELYUM_API_KEY"   # export ELYUM_API_KEY=ek_live_…

Other clients

Windsurf, VS Code, Cline, Zed… any client that takes an mcpServers or servers block. Clients with MCP sign-in need only the URL; the others send the key as a bearer header.

{
  "mcpServers": {
    "elyum-api": {
      "type": "http",
      "url": "https://api.elyum.ai/v1/mcp",
      "headers": { "Authorization": "Bearer ek_live_…" }
    }
  }
}

OAuth

Claude.ai, Claude Desktop, Claude Code, Cursor and most current clients sign in with OAuth. The client registers itself and opens the consent page on elyum.ai, where you choose the scopes, a daily cap in dollars and, inside a business, which client the connection works for. It then holds a token for this server that renders and pays like an API key and renews itself. Connected apps are listed under Keys in your console; disconnecting one there signs it out at once.

API keys

Authorization: Bearer ek_live_…

Keys start with ek_live_ and are made on Keys in your console, where you can also revoke one at once; a revoked key answers 401 auth from then on.

Scopes

Daily spend cap

A key or a connection can carry a daily cap in dollars: what it may spend in a rolling 24 hours. A render that would go over it is refused with 429 key_cap and nothing is held, so a runaway loop stops at the cap, not at an empty balance.

The workflow

1. 1
   get_account

   The balance the API may spend, and the key's scopes and cap.
2. 2
   list_models, get_model

   Pick a model and read what it takes: inputs, settings and their allowed values.
3. 3
   create_generation with dry_run: true

   The price in usage.usd, with nothing held. Free.
4. 4
   create_generation with wait: 55

   Starts the render and holds the answer until it ends; else call get_generation with wait until status is succeeded.
5. 5
   The outputs

   Signed links to the final files, at full quality.

Prices and billing

A render the server starts is a /v1 render in every way: priced like the API, in US dollars, and paid from your API balance, the funds you add on Billing. It is charged when it is delivered and free when it fails or is canceled. It is refused before anything is held when it would cost more than max_usd, the key's daily cap or the balance.

There are no previews and nothing to keep or kill: the file a render returns is the final one.

Answers

Every answer is the /v1 object as it is: a model, a quote, a generation, a file, the account. What REST puts in headers rides the result's _meta: request_id, and idempotent_replayed when an idempotency key answered the first generation.

Errors

A refusal is a tool error that carries the error object (code, message, param, fix, request_id) and a line telling the agent whether to fix the call, retry, or stop and ask you. Every code is in Error codes; these stop the agent: