BananaBanana MCP Server

A remote Model Context Protocol server exposing the full BananaBanana generation stack as agent tools: Google Nano Banana 2 Lite / 2 / Pro images (from $0.03), Veo 3.1 / Fast / Lite video and Gemini Omni Flash with sound (flat $1.00). Your agent sees live prices, confirms costs before expensive runs, and every generation lands in the same history and balance as the website.

• Endpoint: https://bananabanana.pro/api/mcp (Streamable HTTP)
• Auth: Bearer API key from your profile (OAuth 2.1 support is planned)
• Rate limit: 20 tool calls per minute per key; optional per-key daily spend cap
• No MCP client? The same endpoint is plain JSON-RPC over HTTPS — call it from curl, Python or TypeScript without any SDK
• Failed or filtered generations are refunded automatically
• Open docs, server.json and copy-paste client examples in the bananabanana-mcp GitHub repository

1. Create an account and top up the balance. 2. In Profile → MCP API Keys create a key (shown once). 3. Add the server to your client:

Claude Code

claude mcp add --transport http bananabanana https://bananabanana.pro/api/mcp \
  --header "Authorization: Bearer bb_live_YOUR_KEY"

Claude Desktop

Add to claude_desktop_config.json (Settings → Developer → Edit Config); requires Node.js for the mcp-remote bridge:

{
  "mcpServers": {
    "bananabanana": {
      "command": "npx",
      "args": [
        "-y", "mcp-remote", "https://bananabanana.pro/api/mcp",
        "--header", "Authorization: Bearer bb_live_YOUR_KEY"
      ]
    }
  }
}

Cursor

Add to ~/.cursor/mcp.json (global) or .cursor/mcp.json in the project; to keep the key out of the file, use ${env:BB_API_KEY} instead of the literal value:

{
  "mcpServers": {
    "bananabanana": {
      "url": "https://bananabanana.pro/api/mcp",
      "headers": {
        "Authorization": "Bearer bb_live_YOUR_KEY"
      }
    }
  }
}

Full Cursor walkthrough, including quirks and a live-key demo: Generate Images in Cursor.

VS Code / GitHub Copilot

Add to .vscode/mcp.json in the workspace (or your user settings.json); VS Code prompts for the key once and stores it encrypted:

{
  "servers": {
    "bananabanana": {
      "type": "http",
      "url": "https://bananabanana.pro/api/mcp",
      "headers": {
        "Authorization": "Bearer ${input:bb-api-key}"
      }
    }
  },
  "inputs": [
    {
      "type": "promptString",
      "id": "bb-api-key",
      "description": "BananaBanana API key (bb_live_…)",
      "password": true
    }
  ]
}

Full VS Code / Copilot walkthrough, including quirks and a live-key demo: Generate Images in VS Code: Copilot MCP Setup Guide.

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "bananabanana": {
      "serverUrl": "https://bananabanana.pro/api/mcp",
      "headers": {
        "Authorization": "Bearer bb_live_YOUR_KEY"
      }
    }
  }
}

Any other MCP client (raw JSON-RPC)

POST https://bananabanana.pro/api/mcp
Authorization: Bearer bb_live_YOUR_KEY
Content-Type: application/json
Accept: application/json, text/event-stream

{"jsonrpc":"2.0","id":1,"method":"tools/list"}

Prefer a walkthrough? The tutorial Generate Images in Claude Code covers the Claude setup end to end, with four real use cases and their actual costs.

MCP is an open protocol: any client that speaks Streamable HTTP and can attach an Authorization header connects. Current state, verified against each vendor's official docs (last checked July 2026):

Codex example — add to ~/.codex/config.toml and export BB_API_KEY=bb_live_…:

[mcp_servers.bananabanana]
url = "https://bananabanana.pro/api/mcp"
bearer_token_env_var = "BB_API_KEY"

Full Codex walkthrough — CLI, IDE extension and ChatGPT desktop from one config, with quirks and a live-key demo: Codex MCP Setup: Images and Video from One config.toml.

Don't use an MCP client? You don't need one — and you don't need an SDK either. The server is plain JSON-RPC 2.0 over HTTPS: one POST endpoint, a Bearer header, a single JSON response. It is stateless, so there is no session handshake to manage — tools/call works as the very first request from curl, Python, TypeScript or anything else that can send HTTP.

curl

# start an image generation (charges one image, $0.06 on the default model)
curl -s https://bananabanana.pro/api/mcp \
  -H "Authorization: Bearer bb_live_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{
        "name":"generate_image",
        "arguments":{"prompt":"studio photo of a ceramic mug on linen, soft daylight"}}}'

# → result.structuredContent.job_id = "cmxy…"

# fetch the result (long-polls server-side up to 30 s; free)
curl -s https://bananabanana.pro/api/mcp \
  -H "Authorization: Bearer bb_live_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{
        "name":"get_result",
        "arguments":{"job_id":"cmxy…","wait_seconds":30}}}'

Python

Only requests — no MCP library:

import requests

MCP = "https://bananabanana.pro/api/mcp"
HEADERS = {"Authorization": "Bearer bb_live_YOUR_KEY"}

def call(tool, **args):
    r = requests.post(MCP, headers=HEADERS, json={
        "jsonrpc": "2.0", "id": 1, "method": "tools/call",
        "params": {"name": tool, "arguments": args},
    })
    r.raise_for_status()
    return r.json()["result"]["structuredContent"]

job = call("generate_image", prompt="watercolor painting of a lighthouse at dawn")

result = call("get_result", job_id=job["job_id"], wait_seconds=30)
while result["status"] == "processing":
    result = call("get_result", job_id=job["job_id"], wait_seconds=30)

print(result["files"][0]["url"], "cost:", result["cost_charged_usd"])

TypeScript / Node.js

Zero dependencies — built-in fetch (Node 18+, Deno, Bun, browsers):

const MCP = "https://bananabanana.pro/api/mcp";

async function call(tool: string, args: Record<string, unknown>) {
  const res = await fetch(MCP, {
    method: "POST",
    headers: {
      Authorization: `Bearer ${process.env.BB_API_KEY}`,
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      jsonrpc: "2.0", id: 1, method: "tools/call",
      params: { name: tool, arguments: args },
    }),
  });
  const { result } = await res.json();
  return result.structuredContent;
}

const job = await call("generate_image", {
  prompt: "isometric 3D render of a tiny greenhouse at golden hour",
});

let out = await call("get_result", { job_id: job.job_id, wait_seconds: 30 });
while (out.status === "processing") {
  out = await call("get_result", { job_id: job.job_id, wait_seconds: 30 });
}
console.log(out.files[0].url, "cost:", out.cost_charged_usd);

Details that matter when you script it:

• Every tool response carries the same JSON twice: machine-readable result.structuredContent and a text block in result.content — parse whichever is easier.
• Video (and multi-image batches) are a two-step call: the first generate_video returns a quote and charges nothing; repeat it with confirm_cost set to the quoted amount to start.
• Pass an idempotency_key on generation calls — network retries will never double-charge.
• tools/list works without a key, so you can explore the full schemas before creating an account. Tool calls are limited to 20 per minute per key.

• Prices are served live by list_models — the same source the website uses.
• Every video (and multi-image batch) call first returns a quote and charges nothing; the agent repeats the call with confirm_cost to start.
• Every result includes cost_charged_usd and balance_remaining_usd.
• Upstream failures and content-filter rejections are refunded automatically — same policy as the web app.
• Optional idempotency_key guarantees retries never double-charge.

→ generate_video {"prompt": "drone shot over a misty pine forest", "model": "veo-3.1-fast"}
← {"status": "confirmation_required", "quoted_cost_usd": 0.70, ...}
→ generate_video {..., "confirm_cost": 0.70}
← {"job_id": "cmxy…", "status": "processing", "cost_charged_usd": 0.70, "balance_remaining_usd": 12.40}
→ get_result {"job_id": "cmxy…"}
← {"status": "completed", "files": [{"url": "https://…"}], "cost_charged_usd": 0.70}

• Result URLs are signed and valid for 24 hours; the media itself stays in your account — call get_result again for fresh links, or download from the website.
• API keys are stored hashed and shown only once at creation; revoke them any time in your profile.
• Per-key usage log (tool, model, cost, prompt preview) is visible in Profile → MCP API Keys.
• MCP generations appear in the same generation history and analytics as web generations.