AI models & inference

Split text into overlapping chunks for RAG ingestion — by characters or words, with configurable chunk size and overlap. Deterministic (same input → same chunks), returns each chunk with its index and start/end offsets. No API key, no LLM. Pairs with /v1/embeddings for a full chunk→embed pipeline.

Fetch historical orderbook snapshots for a Kalshi market

OpenAI-compatible text embeddings (text-embedding-3-small by default; 3-large and ada-002 supported), up to 64 inputs or 16k chars per call, returned in the standard OpenAI shape. Identical inputs repeated within 10 minutes are served free from cache. Use it for semantic search, clustering and retrieval from any OpenAI SDK by changing base_url.

Get raw book ticker data at microsecond granularity

Convert text to natural speech audio (MP3, WAV or OGG) with neural Piper voices: en-female, en-male (US English) and de-male (German); all voices public domain or CC0, so the audio can be used commercially. Params: text (required, max 2000 characters), voice (default en-female, or picked from language), language (en or de), speed (0.5 to 2, default 1), format (mp3 default, wav, ogg). Returns audio_base64 and duration.

Sentence embeddings (all-MiniLM-L6-v2, 384 dimensions, L2-normalized, multilingual input works best in English) for semantic search, clustering and deduplication, or rerank mode: pass query + documents to get documents sorted by cosine similarity. Runs locally, texts are not stored. Params: texts (1 to 32 strings, max 2000 chars each) for embed mode, or query + documents (up to 64) and top_k for rerank mode.

Translate text between 21 languages: en, de, fr, es, it, pt, nl, pl, sv, cs, el, ru, uk, tr, ar, hi, id, vi, zh, ja, ko. Runs on our own server (Argos Translate), so the text is not sent to a third party. Max 1000 characters per call. source_lang defaults to auto-detect; confidence is the detection confidence. Pairs without a direct model are translated via English.

Put one question to a panel of 3 different AI models and get a single verdict back. Each model answers independently, then a chair synthesizes: verdict, confidence, the points they all agreed on, the dissent, and what would flip the answer. Use it for judgement calls where one model's opinion is not enough: is this safe, is this worth it, which option. No account, no API key, one USDC micropayment per call.

LLM chat completion with Llama 3.3 70B (fast): send messages or a prompt, up to 2048 output tokens.

Turns text into natural speech in English, Spanish, French, Chinese, Japanese or Korean in one or two seconds: Deepgram Aura-2 for English and Spanish, MeloTTS for the rest. Returns a hosted audio URL valid for 24 hours (MP3 for English and Spanish, WAV otherwise) plus its size; add inline=true to also get the base64. Up to 2000 characters per call. You are only charged if the audio is delivered. No API key, no account.

Toxicity detection: classify text as toxic or not toxic (hate speech, harassment, threats, severe offensive language) with a confidence score. Powered by Jev, with an automatic Groq-hosted LLM fallback if Jev is slow or unavailable - under 4.5s guaranteed. Try GET /toxicity/sample.

Claude Sonnet API - pay per call, no API key. OpenAI-format chat completions on anthropic/claude-sonnet-5.5, pinned to Anthropic's own OpenRouter endpoint for reliability. You sign a fixed price computed from your max_tokens; unused tokens are not refunded. 20s server-side timeout - never charged if it fires; set your client timeout to 30s. Try GET /llm/claude-sonnet/sample.

GPT Mini API - pay per call, no API key. OpenAI-format chat completions on openai/gpt-5.4-mini, pinned to OpenAI's own OpenRouter endpoint for reliability. You sign a fixed price computed from your max_tokens; unused tokens are not refunded. 20s server-side timeout with an automatic 2nd-provider retry if the first is slow - never charged if both miss; set your client timeout to 30s. Try GET /llm/gpt-mini/sample.

Infer a basic JSON Schema (types, object properties, array item type) from an example document.

437 models, no account, no API key, pay per call in USDC - OpenAI-compatible chat completions and AI inference gateway (Claude, GPT, Gemini, Llama, Mistral and more in one endpoint). Price: from $0.001, computed from your max_tokens × model rate × 1.10 (you sign a ceiling; unused tokens are not refunded). 20s server-side timeout - never charged if it fires; set your client timeout to 30s. Try GET /v1/chat/completions/sample.

gpt-4.1-mini: OpenAI GPT 4.1 Mini chat completions, paid per call in USDC. Per 1M tokens in/out: Standard (default) 0.4/1.6, Fast 0.7/2.8; pick with service_tier. Plus $0.0005 (Base) or $0.0005 (Solana) per call; the quote charges input plus 10% of max_tokens. Empty answers are free. Rates: https://openai.mm.family/x402/pricing

gpt-6.1-sol (OpenAI's newest model) and 25 more OpenAI chat models (gpt-6, gpt-5.x, gpt-4.1, gpt-4o, o-series), paid per call in USDC. OpenAI token rates on Flex (half price, may answer 429; default where offered), Standard (service_tier default) or Fast (fast), plus $0.0005 (Base) or $0.0005 (Solana) per call; the quote charges input plus 10% of max_tokens. Empty answers are free. One endpoint per model: /x402/v1/models/<key>/chat/completions. Rates: https://openai.mm.family/x402/pricing

OpenAI image generation (gpt-image-2.5, gpt-image-2, gpt-image-1.5, gpt-image-1, gpt-image-1-mini, chatgpt-image-latest), paid per call in USDC. OpenAI list rates per 1M image tokens (1024x1024 low = 196 tokens on gpt-image-2), plus $0.0005 (Base) or $0.0005 (Solana) per call. Send size and quality (auto is refused); omitted means 1024x1024 medium. Returns b64_json. One endpoint per model: /x402/v1/models/<key>/images/generations. Rates: https://openai.mm.family/x402/pricing

Translation for agents handling multilingual DeFi content: converts protocol docs and error messages across languages with tone control and auto source-language detection. Category: inference.

Focused check of a site's /llms.txt and /llms-full.txt: presence, size and llmstxt.org structural validity (an H1 title, a blockquote summary, sections of markdown links). Returns a readiness score plus fixes. Cheap sub-check of the full AEO audit.

Anthropic Messages API over x402 - point the Anthropic SDK (or Claude Code / the Agent SDK) at base_url https://agent402.tools/v1/auto and pay $0.01 per call in USDC, no API key, no signup. Same models, caps and price as this tier's /chat/completions route; any model here is served through the Messages wire (Claude natively, others translated). Depth: `effort` (low..max) on Claude 4.7+, thinking.budget_tokens on older Claude. Up to 16,000 input chars and 1024 output tokens; streaming supported.

OpenAI Responses API over x402 - point the OpenAI SDK's responses.create() (or the OpenAI Agents SDK) at base_url https://agent402.tools/v1 and pay $0.02 per call in USDC, no API key, no signup. Same models, caps and price as this tier's /chat/completions route; any model here is served through the Responses wire.

Generate a text embedding vector using OpenAI text-embedding-3-large (3072 dimensions). Higher accuracy than the small model. Ideal for semantic search, RAG, and clustering. No API key needed; pay per call via x402. Text capped at 32k chars.

Google's native generateContent wire on the auto tier. Send Gemini's own request shape (contents, systemInstruction, generationConfig, function declarations) and get Gemini's own response shape (candidates, usageMetadata) back. Point a Google GenAI SDK at this gateway and pay per request with USDC, no account. Also answers the Google-shaped path /v1beta/models/<model>:generateContent.

Audio URL → finished subtitles in one call: transcribe the audio, emit the transcript as SRT/WebVTT/JSON cues, and report the text statistics - length, reading time, word count. Subtitle pipeline skill pack: one x402 payment runs 3 underlying tools (transcribe, srt-convert, text-stats); partial-success per step.

Anthropic Messages API over x402 - point the Anthropic SDK (or Claude Code / the Agent SDK) at base_url https://agent402.tools/v1 and pay $0.02 per call in USDC, no API key, no signup. Same models, caps and price as this tier's /chat/completions route; any model here is served through the Messages wire (Claude natively, others translated). Depth: `effort` (low..max) on Claude 4.7+, thinking.budget_tokens on older Claude. Up to 32,000 input chars and 2048 output tokens; streaming supported.

Google's native generateContent wire on the base tier. Send Gemini's own request shape (contents, systemInstruction, generationConfig, function declarations) and get Gemini's own response shape (candidates, usageMetadata) back. Point a Google GenAI SDK at this gateway and pay per request with USDC, no account. Also answers the Google-shaped path /v1beta/models/<model>:generateContent.

NHTSA safety recalls for a vehicle by make/model/year: campaign number, affected component, summary, consequence, and remedy. Live gov data, no key.

Google's native generateContent wire on the nano tier. Send Gemini's own request shape (contents, systemInstruction, generationConfig, function declarations) and get Gemini's own response shape (candidates, usageMetadata) back. Point a Google GenAI SDK at this gateway and pay per request with USDC, no account. Also answers the Google-shaped path /v1beta/models/<model>:generateContent.

Google's generateContent wire priced from the request itself: the 402 quotes this exact body, from $0.001, with upto buyers settling actual usage. Use it when an agent built on the Gemini SDK needs a supported model paid per call without an API key.