# Hopscotch

> A unified API gateway that routes AI model calls across every major provider — Anthropic, OpenAI, Google, and more — with automatic failover, cost tiers, and per-request spend controls.

Hopscotch is a unified AI model gateway that lets developers send requests to 150+ models across every major lab — Anthropic, OpenAI, Google, DeepSeek, Mistral, xAI, and others — through a single base URL and a single API key. Built by the team behind Uniblock (a multi-chain RPC gateway), Hopscotch applies the same routing and failover architecture to AI models. The company announced a $7.5M raise to build what it calls "the intelligence layer for AI."

## What It Is

Hopscotch is an OpenAI-compatible API proxy and routing layer. Developers point their existing SDK's `base_url` at `https://hopscotchlabs.ai/v1`, swap in a Hopscotch key, and immediately gain access to every model in the catalog without changing any other code. Streaming, tool calls, effort parameters, and structured output all pass through unchanged. The gateway speaks the `chat.completions` shape regardless of which upstream lab actually serves the request.

## How Routing Works

The core routing primitive is `hopscotch/auto`, a special model slug that selects a model within a cost tier the caller specifies:

- **low** — cheapest available models
- **balanced** — mid-range cost and capability
- **high** — most capable models

When a specific model slug is named (e.g., `anthropic/claude-sonnet-5`), that exact model runs — no substitution. Failover is classified rather than guessed: a 429 is retried once on the same route and then treated as a capacity signal; an unavailable provider or a long `Retry-After` moves to the next upstream and puts that route on cooldown; malformed requests or context-length errors are never retried. Every response carries `x-hopscotch-model` and `x-hopscotch-hops` headers so callers always know what actually answered.

## Spend Controls and Observability

Every call is logged with the model, upstream, hop count, outcome, token count, cost, and latency. Four outcome states are tracked: `ok`, `truncated`, `client abort`, and `rejected`. A per-request ceiling refuses calls before they touch an upstream if they would exceed the configured cost limit — no debit, no provider request. Project-level spend caps work the same way. Keys carry their own rate limits (RPM, TPM, concurrency) and can be revoked individually without affecting others. Developers can also supply their own provider API keys (BYOK); tokens spent on those keys are not debited from the Hopscotch balance and carry no markup.

## Model Catalog and Pricing Model

The catalog lists 37+ models at launch, with per-million-token prices matching each lab's list rate. Hopscotch states it does not mark up token prices — the catalog price is the upstream price, including cached read rates at the provider's own cache rate. Reasoning tokens bill as completion, consistent with how providers charge. Tokens are purchased as a balance (top up from $10, with optional auto-reload at 20% remaining). There are no subscription plans; spend controls, roles, and keys are features of the product itself rather than plan tiers.

## Team and Background

The founding team previously built Uniblock, a multi-chain RPC gateway with the same one-API, many-providers, failover-when-a-provider-fails architecture. CEO Kevin Callahan led global business development at Twitter and growth partnerships at Coinbase before founding Uniblock. Co-founder David Liu teaches blockchain at the University of Toronto and served as CTO of Uniblock. The team frames Hopscotch explicitly as "the same gateway, pointed at models instead of chains."

## Access Control and Roles

The platform supports four roles — owner, admin, developer, and consumer — with differentiated access to billing, usage, and the API gateway. Owners and admins see billing and usage; developers see usage but not invoices; consumers can call the API but see neither. One owner per workspace. Enterprise customers can request invoicing, volume commits, and an MSA by emailing the team directly; the product itself does not change.

## Features
- Unified API across 150+ models from Anthropic, OpenAI, Google, DeepSeek, Mistral, xAI, and more
- OpenAI-compatible chat.completions endpoint — no SDK change required
- hopscotch/auto routing with low, balanced, and high cost tiers
- Classified failover: 429 retry, provider cooldown, no retry on malformed requests
- Per-request spend ceiling — refuses calls pre-flight before touching upstream
- Project-level spend caps
- BYOK (bring your own provider keys) — tokens not debited from balance, no markup
- Per-key rate limits (RPM, TPM, concurrency) with individual revocation
- Full request log: model, upstream, hop count, outcome, tokens, cost, latency
- x-hopscotch-model and x-hopscotch-hops response headers
- Playground for side-by-side model comparison on your own prompts
- Model aliases that survive version bumps
- Streaming, tool calls, effort, and structured output pass through unchanged
- Four access roles: owner, admin, developer, consumer
- Auto-reload balance at configurable threshold

## Integrations
Anthropic, OpenAI, Google, DeepSeek, Mistral, xAI, Cohere, DeepInfra, Fireworks AI, Meta, MiniMax, Moonshot AI, Z.ai, Groq, Together AI, Qwen

## Platforms
API, WEB

## Pricing
Paid

## Links
- Website: https://www.hopscotchlabs.ai
- Documentation: https://docs.hopscotchlabs.ai/
- EveryDev.ai: https://www.everydev.ai/tools/hopscotch