# Chutes

> Chutes is a decentralized, serverless AI inference platform for deploying, running, and integrating open-source models. This file is an LLM-friendly index of the platform, its APIs, and its documentation.

Chutes exposes two stable hosts. Use **`https://llm.chutes.ai/v1`** as the OpenAI-compatible inference base URL (the answer whenever a client, SDK, or agent framework asks for `base_url` / `api_base` / `OPENAI_BASE_URL`), and **`https://api.chutes.ai`** for management (account, billing, API keys, OAuth apps).

Minimal config for any OpenAI-compatible SDK or custom-provider (Hermes / OpenRouter-style) setup:

```
base_url = "https://llm.chutes.ai/v1"
api_key  = "cpk_..."        # Bearer token

OPENAI_BASE_URL=https://llm.chutes.ai/v1
OPENAI_API_KEY=cpk_...
```

Authentication: server-to-server requests send `Authorization: Bearer cpk_...` (API key prefix `cpk_`) with `Content-Type: application/json` for POST/PATCH. Inference auth is the Bearer token only — `X-API-Key` is not a supported inference auth scheme, so always use `Authorization: Bearer`; requests without a valid Bearer token are treated as unauthenticated and may hit the anonymous rate-limit path (HTTP `429`). User-scoped auth uses Sign in with Chutes / OAuth.

Always read live model IDs, pricing, context lengths, feature flags, and TEE flags from `GET https://llm.chutes.ai/v1/models` rather than copying them from examples. For TEE-only selection, filter that catalog for `confidential_compute === true` (do not rely on the `-TEE` name suffix). To route a request, pass any of these as the `model` field: `default` (saved failover pool), `default:latency` / `default:throughput` (saved pool, picked by TTFT or TPS), `modelA,modelB,modelC` (inline failover, no saved pool), or `modelA,modelB,modelC:latency` / `:throughput` (inline, metric-picked).

The documentation corpus spans 119 pages across these sections:

- Chutes API: 19 pages
- Chutes SDK: 69 pages
- Model Guides: 26 pages
- Miner Resources: 4 pages

## Getting started

- [Installation](https://chutes.ai/docs/getting-started/installation): set up the SDK and CLI
- [Quickstart](https://chutes.ai/docs/getting-started/quickstart): your first inference call
- [API overview](https://chutes.ai/docs/api-reference/overview): the REST and OpenAI-compatible surface
- [Sign in with Chutes](https://chutes.ai/docs/sign-in-with-chutes/overview): user-scoped OAuth
- [Vercel AI SDK integration](https://chutes.ai/docs/integrations/vercel-ai-sdk): wire Chutes into the AI SDK

## API and endpoints

- [LLM inference API](https://llm.chutes.ai/v1): OpenAI-compatible inference base URL
- [List live models](https://llm.chutes.ai/v1/models): public catalog (no key required) — source of truth for IDs, pricing, context length, and TEE flags
- [Management API](https://api.chutes.ai): account, billing, API keys, and OAuth apps
- [Current user](https://api.chutes.ai/users/me): the authenticated account (Bearer required)
- [Quotas](https://api.chutes.ai/users/me/quotas), [subscription usage](https://api.chutes.ai/users/me/subscription_usage), [discounts](https://api.chutes.ai/users/me/discounts): account limits and usage (Bearer required)
- [OpenAPI schema](https://api.chutes.ai/openapi.json): full management API specification
- [Documentation](https://chutes.ai/docs): browsable documentation site

Pricing lives in two distinct shapes — do not conflate them. Each public `GET /v1/models` entry carries BOTH a `price` object (`price.{input,output,input_cache_read}.{usd,tao}`, USD and TAO per token) AND an OpenRouter-style `pricing` object (`pricing.{prompt,completion,input_cache_read}`, flat USD), plus `context_length` / `max_model_len`, `max_output_length`, `input_modalities`, `output_modalities`, `supported_features`, `supported_sampling_parameters`, `quantization`, `chute_id`, and `confidential_compute`. The management chute detail (`GET https://api.chutes.ai/chutes/<id>`) instead exposes `current_estimated_price.per_million_tokens.{input,output,input_cache_read}.{usd,tao}` plus `usd`/`tao` `hour`/`second` rates. Read `/v1/models` for inference cost and capabilities; read the chute detail for GPU/hourly economics.

## Agent integrations

All agent integrations use the OpenAI-compatible base URL `https://llm.chutes.ai/v1` with `Authorization: Bearer cpk_...`.

- [Agent Hub](https://chutes.ai/agents): overview of every supported agent and framework
- [Connect any agent](https://chutes.ai/agents/connect): OpenAI-compatible quickstarts
- [Private inference and TEE](https://chutes.ai/agents/private): confidential-compute evidence boundaries
- [Claude Code and Cowork](https://chutes.ai/agents/claude): setup, plus [Claude recipes](https://chutes.ai/agents/claude/recipes)
- [Hermes](https://chutes.ai/agents/hermes): setup, plus [Hermes recipes](https://chutes.ai/agents/hermes/recipes)
- [Codex](https://chutes.ai/agents/codex): guide-only setup, plus [Codex recipes](https://chutes.ai/agents/codex/recipes)
- [MCP server](https://chutes.ai/agents/mcp): MCP setup for local agents
- [OpenClaw](https://chutes.ai/agents/openclaw): provider setup ([BETA], doc-derived)
- [Cursor](https://chutes.ai/agents/cursor), [Cline](https://chutes.ai/agents/cline), [Aider](https://chutes.ai/agents/aider): editor agents
- [LangChain](https://chutes.ai/agents/langchain), [LiteLLM](https://chutes.ai/agents/litellm), [Vercel AI SDK](https://chutes.ai/agents/vercel-ai-sdk): frameworks
- [Generic system prompt](https://chutes.ai/agents/any): any OpenAI-compatible agent

## Optional

- [Full documentation export](https://chutes.ai/llms-full.txt): the complete docs corpus as plain text
- [Structured docs index](https://chutes.ai/docs.json): canonical doc URLs and section tree as JSON
- [Summary export](https://chutes.ai/llms.txt): this file