Hopscotch AI

Product Introduction

1. Definition: Hopscotch AI is a unified API gateway and intelligent routing platform for large language models (LLMs). It acts as a single integration point, providing developers with programmatic access to over 150 AI models from providers like OpenAI (GPT-5.6), Anthropic (Claude Sonnet 5), Google (Gemini 3.6 Flash), DeepSeek, and Meta.
2. Core Value Proposition: It eliminates the complexity of managing multiple AI provider integrations, separate API keys, and disparate billing systems. Its primary value is offering cost-effective, reliable, and simplified LLM access through features like automatic failover, cost-tiered routing, and unified observability, all without token markup.

Main Features

1. Unified API Endpoint: Hopscotch provides a single, OpenAI-compatible API base URL (https://hopscotchlabs.ai/v1). Developers can point their existing OpenAI SDK client to this endpoint and call any supported model by specifying its unique slug (e.g., anthropic/claude-sonnet-5). This feature ensures backward compatibility and eliminates the need for multiple SDKs.
2. Intelligent Model Routing & Auto Tier: The platform's core intelligence is the hopscotch/auto model. When specified, it automatically selects the most appropriate model from a user-defined cost-performance tier: Low, Balanced, or High. This dynamic routing optimizes for cost-efficiency or capability without manual model selection. Users can also define explicit, ordered fallback chains for specific models.
3. Granular Cost Controls & Observability: Hopscotch implements robust financial safeguards. Users can set per-request cost ceilings and project spending caps. Requests that would exceed these limits are rejected pre-flight, preventing runaway costs from agent loops. Every API call is logged with detailed metadata: the final serving model, all attempted upstreams ("hops"), token usage, cost, latency, and outcome, providing complete auditability in one dashboard.
4. Bring-Your-Own-Key (BYOK) & Transparent Pricing: Users can attach their own provider API keys (e.g., an OpenAI key) to their Hopscotch project. These keys are used first, and tokens consumed via them are not debited from the Hopscotch balance and incur no platform markup. The platform charges the upstream provider's list price directly, passing through cached response discounts and charging reasoning tokens as completions.

Problems Solved

1. Pain Point: Developer fragmentation and operational overhead from integrating and maintaining connections to numerous, differently-structured LLM APIs, each with its own authentication, error handling, and rate limits.
2. Pain Point: Unpredictable costs and lack of spend control when experimenting with or deploying multiple LLMs, risking budget overruns from inefficient model choices or faulty agentic loops.
3. Target Audience: AI Engineers and ML Developers building production applications who need reliability and cost-optimization; Startup CTOs managing tight R&D budgets for AI features; Enterprise DevOps Teams requiring centralized logging, security, and compliance across AI usage.
4. Use Cases: A/B Testing LLMs: Using the Playground to send identical prompts to different models (e.g., Claude 3.5 Sonnet vs. GPT-4o) for direct, cost-aware comparison. Building Resilient AI Agents: Configuring fallback routes (e.g., primary: GPT-4, fallback: Claude Sonnet) to ensure uptime during provider outages. Unifying AI Analytics: Gaining a single pane of glass for usage, performance, and spend metrics across all AI providers used by a development team.

Unique Advantages

1. Differentiation: Unlike aggregators that charge a percentage markup, Hopscotch operates on a pass-through pricing model. Unlike simple proxy services, it offers sophisticated, configurable routing logic and financial guardrails. It differs from vendor lock-in platforms by supporting BYOK and offering a unified interface without altering request/response payloads.
2. Key Innovation: The classified failover system is a technical differentiator. It intelligently categorizes upstream errors (e.g., 429, 503, context length errors) and applies appropriate retry, failover, or immediate failure policies, preventing wasteful retries on unrecoverable errors and optimizing for success and latency.

Frequently Asked Questions (FAQ)

1. Does Hopscotch AI mark up the cost of AI models? No, Hopscotch AI employs a transparent, pass-through pricing model. You pay the exact input and output token rates charged by the upstream provider (e.g., OpenAI, Anthropic). When using your own provider keys (BYOK), you pay the provider directly with no Hopscotch fees.
2. How does Hopscotch prevent an AI agent from causing unlimited spending? Hopscotch implements two key financial controls: a per-request cost ceiling and a project spending cap. If an API call is predicted to exceed either limit, it is rejected before any upstream call is made, preventing token spend and providing an immediate error for your application to handle.
3. Can I force Hopscotch to always use a specific model version? Yes. When you specify an exact model slug (e.g., openai/gpt-4o-2024-08-06), that model will always run. The platform also provides aliases (e.g., openai/gpt-4o) that point to a recommended version, but pinning is explicitly controlled by the developer's choice of slug.
4. What happens if my chosen AI provider is down or rate-limited? Hopscotch's intelligent routing can be configured with fallback models. For the hopscotch/auto tier, it handles this automatically. For named models, you can define a sequential fallback chain. The system classifies errors—retrying rate limits (429) briefly, while quickly failing over for outages—minimizing downtime.
5. Do I need to rewrite my code to use Hopscotch AI? Typically, only one line of code needs changing: the base URL in your OpenAI SDK client. The API is fully compatible with the OpenAI format, supporting streaming, function calling (tools), structured outputs, and other advanced features without modification.