Banana Peel

Banana Peel is a smart routing API for browser agents that sits in front of 50+ single runners and exposes them as 250+ routable combinations behind one key and one bill. It targets developers building AI agents that need to automate web tasks, offering parallel fallback, per-domain learning, and a unified response format so agents don't have to manage multiple provider integrations.

What It Is

Banana Peel is an API-first routing layer in the browser automation infrastructure category. Instead of picking one browser runner and hardcoding it, developers send a single POST /api/v1/responses request with a routing mode — smart, a named runner, a category, or a short list — and Banana Peel selects the runner most likely to succeed on the target domain. The API is OpenAI-compatible in structure, and the platform provides drop-in compatibility wrappers for Browserbase, Steel, Browser Use, Skyvern, and Hyperbrowser so existing code can migrate by swapping an API key and base URL.

How Smart Routing Works

The routing layer maintains a live per-domain scoreboard updated by every run that flows through the platform. When routing: smart is selected, the router starts with the runner that last succeeded on the target domain — no cold guesses. If a run is blocked, Banana Peel does not queue retries; instead it fans out to two or three runners in parallel and takes the first success. Every attempt — including failed and parallel fallback attempts — is itemized on the response receipt, which includes screenshots, artifacts, session details, and per-attempt costs. The site publishes a live benchmark drawn from real traffic at /benchmarks.

Runner Network and Integrations

The platform covers 50+ single runners spanning hosted agents, browser infrastructure providers, and open-source frameworks. Named runners visible in the source include Browserbase, Steel, Skyvern, Browser Use, Hyperbrowser, Browserless, Notte, Airtop, Anchor Browser, Apify, Firecrawl, Lightpanda, TinyFish, Kernel, and Deck, among others. Open-source frameworks like Playwright and Browser Use OSS are also supported; runs on those debit only metered compute and model tokens rather than a vendor price. The platform also exposes an official MCP server for use with Cursor and Claude, an agent skill endpoint, an llms.txt condensed for context windows, and a full OpenAPI spec.

Deployment Model and Setup Path

Access is entirely API-based — developers get an API key, set BANANA_PEEL_API_KEY, and call POST /api/v1/responses. The platform provides a CLI (npx -y @banana-peel/cli init --agent --json) designed for non-human agent onboarding so agents can mint their own key without a human in the loop. A console at /runs shows run history and session replay. The platform also supports webhooks with signed event delivery and a documented retry ladder, and a credential vault with MFA relay for authenticated web tasks.

Tradeoffs to Know

Every runner attempt that executes is billed — including failed, blocked, and parallel smart-routing fallback attempts. The routing: learning and routing: learning-all modes exist to calibrate the router on a domain or task by sweeping the live runner pool; the site notes these are paid calibration sweeps and should not be used for every production job. Output normalization — reshaping raw run output to a custom JSON schema — is an optional add-on charged per run on top of runner cost and the platform fee. Custom data retention beyond the default 7-day window, SSO, bring-your-own-key (BYOK), regional execution controls, and contractual SLAs are available under Enterprise contract terms.