Banana Peel MCP server

Install & run

Package: @banana-peel/mcp. Transport is stdio. No repo clone required. Banana Peel detects your agent environment and configures itself when you run npx -y @banana-peel/cli init --agent --json.

export BANANA_PEEL_BASE=https://bananapeel.com
export BANANA_PEEL_API_KEY=bp_live_…

npx -y @banana-peel/mcp

# optional global install
npm install -g @banana-peel/mcp
bananapeel-mcp

Auth: BANANA_PEEL_API_KEY (alias BANANAPEEL_API_KEY). Optional base: BANANA_PEEL_BASE (default https://bananapeel.com).

Cursor

User config: ~/.cursor/mcp.json. Project: .cursor/mcp.json when that folder already exists. The CLI writes this automatically during agent-native signup.

{
  "mcpServers": {
    "bananapeel": {
      "command": "npx",
      "args": ["-y", "@banana-peel/mcp"],
      "env": {
        "BANANA_PEEL_API_KEY": "bp_live_…"
      }
    }
  }
}

Claude Desktop

Add to your Claude Desktop claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json). Same npx -y @banana-peel/mcp shape. The CLI writes this when Claude Desktop is installed.

{
  "mcpServers": {
    "bananapeel": {
      "command": "npx",
      "args": ["-y", "@banana-peel/mcp"],
      "env": {
        "BANANA_PEEL_API_KEY": "bp_live_…"
      }
    }
  }
}

Other environments the CLI configures

npx -y @banana-peel/cli init --agent --json detects which of these are present and merges Banana Peel MCP (a failed write never fails init):

• Claude Code — user ~/.claude.json (mcpServers) and skill ~/.claude/skills/bananapeel/SKILL.md
• Codex — ~/.codex/config.toml ([mcp_servers.bananapeel])
• VS Code — user mcp.json under the VS Code profile, and workspace .vscode/mcp.json (servers key, not mcpServers) when .vscode already exists
• Windsurf — ~/.codeium/windsurf/mcp_config.json

Tools

• create_response — POST /api/v1/responses
• get_response — GET /api/v1/responses/:id — Optional wait: true polls until completed, failed, or requires_action.
• submit_input — POST /api/v1/responses/:id/input
• cancel_response — POST /api/v1/responses/:id/cancel
• create_long_task — POST /api/v1/long-tasks
• get_long_task — GET /api/v1/long-tasks/:id
• cancel_long_task — POST /api/v1/long-tasks/:id/cancel
• list_long_task_runs — GET /api/v1/long-tasks/:id/runs
• list_runs — GET /api/v1/runs
• list_rankings — GET /api/v1/rankings
• list_capabilities — GET /api/v1/capabilities

Tools map one-to-one onto the public API — see Create a response and OpenAPI.

Output normalization (+$0.10)

On create_response, pass normalize: true with an output_schema (JSON Schema or property map) to reshape the runner result after the run. Flat $0.10 on that run. Build schemas in Console → Normalize. Full HTTP example: Create a response → Output normalization.

MFA / human input

When a target site asks for MFA, 2FA, or an OTP, the run pauses immediately and asks the user for the code. It does not keep trying other runners.

CAPTCHA and other bot-detection gates (reCAPTCHA, hCaptcha, Turnstile, DataDome, Cloudflare challenge, “prove you are human”) are solved, re-routed, or end blocked with reason bot_challenge. They never become a human text prompt.

create_response waits until the run completes or the site asks for a code. The tool result then includes human_action_required: true and an mfa object (channel, message, submit path) so you do not have to parse raw JSON. ASK THE HUMAN IN CHAT NOW for the site MFA/OTP code. Call submit_input with their code. Do not retry the task or pick another runner.

1. Call create_response (default wait: true).
2. If the result has human_action_required or status: requires_action, ask the human in chat immediately.
3. submit_input with code or answer — Submit the code with POST /api/v1/responses/{id}/input { "code": "123456" } (or { "answer": "…" } for text/confirm). MCP: submit_input({ id, code }). Then keep polling until completed or failed.
4. get_response with wait: true until terminal. Do not start a new run.

# create_response waits until completed | failed | requires_action
create_response({
  input: "Log into the site and extract the invoice total",
  url: "https://example.com/login",
  routing: "smart"
})
# → human_action_required: true
#   status: "requires_action"
#   mfa: { type: "mfa", channel: "sms", message: "…", submit: { method: "POST", path: "/api/v1/responses/resp_…/input" } }
#   required_action: { type: "submit_input", submit_input: { kind: "otp", prompt: "…" } }

# ASK THE HUMAN IN CHAT NOW, then:
submit_input({ id: "resp_…", code: "123456" })
# or: submit_input({ id: "resp_…", answer: "yes" })

# poll to terminal
get_response({ id: "resp_…", wait: true })

If the vault credential (or request totp_secret) includes a base32 TOTP seed, OTP asks are answered in-process and the run does not pause. Username/password alone is not enough — the run still pauses for the code. Full guide: MFA / OTP.

Also see

• For AI agents
• /llms.txt
• /agent-skill
• Quickstart