Nebius Hackathon — Prerequisites Checklist

1✅ Get your account provisioned

CRAFT's hackathon environment uses your existing Google or Microsoft identity — there's no separate password to create. Provisioning is allow-list based, so do this first; it's the only step with a lead time you don't control.

• Submit your Google or Microsoft account email via the Enterprise Agents Hackathon — By Emergence, in partnership with Nebius (NYC) Luma page
• Wait for your confirmation email. Allow-list reviews run in batches — if you haven't heard back within a few business days, ask in #hack
• Do not try to sign in before you get the confirmation — you'll hit a 403 and have to retry later

2🔑 Find your Project ID

Every API call to CRAFT is scoped to a project via the X-Project-ID header. You get this once, from the browser, after your first sign-in.

• Go to https://nebius.emergence.ai and sign in with Sign in with Google or Sign in with Microsoft, using the account you registered with
• Once logged in, your first project is auto-selected and appended to the URL: ?projectId=<uuid>
• Copy that UUID — it starts with e.g. edb5c0bf-5407-..., not em_hk_... or any other prefix. You'll paste it into every MCP client config below.

3🛠️ Install an MCP client

Pick one to start (you can add more later). All of them speak the same protocol — Streamable HTTP + OAuth 2.1 (Authorization Code + PKCE) — so the differences below are purely in where you paste the config, not what the config means.

What every config below has in common:

claude mcp add-json Craft '{
  "type": "http",
  "url": "https://nebius.emergence.ai/mcp",
  "headers": {
    "X-Project-ID": "<your project UUID>"
  },
  "oauth": {
    "clientId": "em-runtime-mcp",
    "authServerMetadataUrl": "https://runtime.prod.emergence.ai/keycloak/realms/hub/.well-known/openid-configuration",
    "callbackPort": 9876,
    "scopes": "openid profile email organization"
  }
}'

{
  "servers": {
    "Craft": {
      "url": "https://nebius.emergence.ai/mcp",
      "type": "http",
      "headers": {
        "X-Project-ID": "<your project UUID>"
      },
      "oauth": {
        "clientId": "em-runtime-mcp",
        "callbackPort": 9876,
        "authServerMetadataUrl": "https://runtime.prod.emergence.ai/keycloak/realms/hub/.well-known/openid-configuration",
        "scopes": "openid profile email organization"
      }
    }
  },
  "inputs": []
}

{
  "mcpServers": {
    "Craft": {
      "type": "http",
      "url": "https://nebius.emergence.ai/mcp",
      "headers": {
        "X-Project-ID": "<your project UUID>"
      },
      "auth": {
        "CLIENT_ID": "em-runtime-mcp",
        "scopes": ["openid", "profile", "email", "organization"]
      }
    }
  }
}

{
  "mcpServers": {
    "Craft": {
      "httpUrl": "https://nebius.emergence.ai/mcp",
      "headers": {
        "X-Project-ID": "<your project UUID>"
      },
      "oauth": {
        "enabled": true,
        "clientId": "em-runtime-mcp",
        "authorizationUrl": "https://runtime.prod.emergence.ai/keycloak/realms/hub/protocol/openid-connect/auth",
        "tokenUrl": "https://runtime.prod.emergence.ai/keycloak/realms/hub/protocol/openid-connect/token",
        "redirectUri": "http://localhost:9876/oauth/callback",
        "scopes": ["openid", "profile", "email", "organization"]
      }
    }
  }
}

mcp_oauth_callback_port = 9876

[mcp_servers.Craft]
url = "https://nebius.emergence.ai/mcp"
scopes = ["openid", "profile", "email", "organization"]

[mcp_servers.Craft.http_headers]
X-Project-ID = "<your project UUID>"

[mcp_servers.Craft.oauth]
client_id = "em-runtime-mcp"

4🔍 Verify your connection

Before the event, confirm end-to-end connectivity — don't wait until hackathon morning:

• Call the hello_world tool (pure sanity check, no upstream dependencies) — if this fails, it's an auth/connectivity problem, not a data problem
• Call list_databases and confirm you see catalog entries — this exercises your X-Project-ID header. If hello_world works but list_databases errors or returns empty, your header likely isn't reaching the server
• If both work, you're fully connected. 401 → re-run OAuth login (cached tokens can go stale). 403 or empty results with no error → project isn't provisioned yet, ping #hack

5🧰 Available tools (current, as of this writing)

Tools show up in your client as <tool-name> or mcp__<server-name>__<tool-name> depending on the client. There are no spider2_/codegen_/dx_/plotly_ prefixes in the shipped server — that naming was from an earlier design and did not ship this way.

6🤖 Choose your agent framework

Any framework works — Claude Agent SDK, Google ADK, LangGraph, Pydantic AI, CrewAI, or hand-rolled. The MCP server doesn't care what's calling it. What differs is how each framework wires in an LLM (Step 7):

7🔐 Get LLM access — Nebius Token Factory (co-sponsor credits) or your own key

Your agent's reasoning LLM is separate from CRAFT's own SQL-generation LLM (generate_sql uses its own backend — you don't need an LLM key just to call CRAFT tools). You only need an LLM for your agent's own orchestration/reasoning loop.

• Request credits via the Nebius Builder Program: dev.nebius.com/builders — apply (describe your background/use case), then claim credits. Confirmed as of 2026-07-09: $50 Nebius Token Factory credit + $50 Nebius AI Cloud credit + $25 Tavily credit + a $1 Nebius Academy certification offer. General early-preview builder program (not hackathon-exclusive) — apply as soon as you know you're participating
• Once approved, get your API key from the Nebius Token Factory console
• Nebius Token Factory is OpenAI-API-compatible — base_url = https://api.tokenfactory.nebius.com/v1/, standard Authorization: Bearer <key>. Plugs directly into LangGraph, Pydantic AI, Google ADK, or any OpenAI-client-based agent
• Recommended model for agentic/tool-calling work: nvidia/nemotron-3-super-120b-a12b — 120B params (12B active, hybrid MoE), up to 1M context, native function calling, OpenAI-compatible
• The $25 Tavily credit is a bonus for this hackathon specifically — the Setup Guide calls out Tavily for web search (CRAFT doesn't proxy it; call it directly for anything outside data intelligence). Grab that key from the same builder-program claim flow
• If building with Claude Agent SDK specifically, Nebius Token Factory does not host Anthropic models — you'll need your own Anthropic API key for that framework's reasoning loop. The Nebius builder credits cover the OpenAI-compatible-endpoint frameworks only
• Standard (non-credit) Token Factory pricing for reference if you exceed the $50: ~$0.30/1M input, ~$0.90/1M output tokens for Nemotron 3 Super

8📊 Know the data you'll be working with — Spider 2.0 subset (9 databases, ~37 GB)

CRAFT pre-loads 9 enterprise databases from the Spider 2.0 benchmark (ICLR 2025 Oral), grouped into 5 challenge scenarios. You don't need to know SQL — generate_sql handles it, including the hardest patterns (LATERAL FLATTEN, multi-schema JOINs, nested VARIANT extraction). Full challenge briefs live on their own Confluence pages.

9⚠️ Known gaps

• MCP is the only interface. A2A multi-agent endpoints, a Python SDK (craft-sdk), and a REST API are planned — don't build a submission that depends on them
• Nemotron-3 Super integration — CRAFT's own generate_sql currently runs on GCP Vertex AI, not Nebius Token Factory. This is a separate, in-progress workstream — it doesn't affect your ability to use Nebius Token Factory for your own agent's LLM (Step 7)