Claude Code / any header-capable MCP client — works today

# the assistant profile: the curated process surface
# + a propose-then-confirm gate on anything that spends
claude mcp add --transport http spoton \
  "https://api-production-eb38.up.railway.app/mcp?profile=assistant" \
  --header "Authorization: Bearer sk_spoton_…"

# integrators: the full ~890-tool surface, one-shot
{ "url": "https://api-production-eb38.up.railway.app/mcp",
  "headers": { "Authorization": "Bearer sk_spoton_…" } }

Works with every MCP client that can send an Authorization header — Claude Code, agent SDKs, IDE clients. On the assistant profile, tools that spend, publish or delete first answer a plain-language proposal with a confirm token; nothing runs until the person agrees and the model calls again with that token. Reads are instant. Stateless transport: no session pinning, safe across our deploys.

Claude.ai & ChatGPT connectors — honest status

Claude.ai and Claude Desktop can send an API-key request header on custom connectors where Anthropic's request-headers beta is enabled on your account — point the connector at /mcp?profile=assistant with Authorization: Bearer sk_spoton_…. Outside that beta, and on ChatGPT (whose connectors authenticate with OAuth 2.1 only), a SpotOn connection needs our OAuth-for-MCP flow, which hasn't shipped yet — that's the honest state, not a configuration you're missing. The endpoint already speaks the standard transport, so when OAuth-for-MCP lands, nothing about your setup changes.

Try the journeys from a connected client: spoton_create_and_publish · spoton_make_variants_and_pick_best · spoton_pitch_client. Actions that need a human sign-off hold at review and say so — the assistant gets the share link, not a bypass.

Every tool carries the four MCP hints (read-only, destructive, idempotent, open-world), derived from the route it calls, so a client can approve reads on its own. An unknown argument is an error that names the accepted ones. spoton_batch runs several tools in order with stop_on_error, and refuses a destructive step unless you allow it. Send a progress token on a tool that starts a spot's run and it reports progress for up to two minutes; cancel that call and the run stops spending.

1 · Mint a scoped key

POST /admin/api-keys
{
  "name": "my app",
  "scope": ["ttsApi"],
  "budgets": { "maxTokensPerDay": 5000 }
}
# → { id, apiKey: "sk_spoton_…" }
#   (the plaintext is shown ONCE)

2 · Synthesize

POST /tts/synthesize
Authorization: Bearer sk_spoton_…
{ "text": "Hello from SpotOn",
  "voiceId": "…" }
# → hosted URL + <audio> embed
#   + a presigned download

3 · Embed the result

<audio controls preload="none"
  src="https://…/tts/public/<token>"
  title="Made with SpotOn"></audio>