OmniRail

The OmniRail platform and its token are now live. Your agent can write code. Now it can also render images, edit them, cut videos and ask any other model for a second opinion. OmniRail gives it all of them: 429 LLMs, image, video and editing models behind one key and one MCP server. It pays autonomously in USDG. omnirail.org CA (Robinhood chain) : 0xa4ddf100e3e0d3a9dd897e89d025622b794f0c08

Webhooks on OmniRail, for builders 🔔 Pass webhook_url on POST /v1/videos, /v1/jobs or /v1/trade/submit (or in the MCP tools). When the job ends we POST one event: video.completed, job.completed, trade.completed (or .failed), with the same object GET returns. Signed with the Standard Webhooks format, so any existing verifier works. Retried for about 43 hours. Only public https addresses are called. Code to check the signature in Node and Python is in the docs: omnirail.org/docs#webhooks

Already using the OpenAI SDK? Change one line. baseURL: "omnirail.org/api/v1" That's it. Your code now reaches 358 chat models: Claude, GPT, Gemini, DeepSeek, Qwen and the rest. Switching models means changing one string. Streaming, tools, JSON output, reasoning and images work the same way. The same key also makes images, video, voice and embeddings. No new SDK. No card. You pay per call from a balance topped up in crypto. omnirail.org/docs

OmniRail shipped 38 updates in its first 5 weeks 🚆 Every one is dated on our changelog, from launch on September 30: → Apps: Chat, the Playground, an Android app → Agents: an Agent Skill, Claude Code on any model, x402, MCP → Control: key limits, scoped keys, webhooks → Models: Compare, Jev decision models, uncensored models → Trade: swaps on Robinhood Chain, routed through Uniswap and its Aggregator Hooks All of it behind one key, paid in crypto. Follow the next ones, there's an RSS feed → omnirail.org/changelog

Your coding agent, your desktop assistant, your custom agent: give it all of OmniRail in one line of config. The OmniRail MCP server exposes the whole catalog as tools. Your agent can list models, generate an image, render a video, transcribe audio or ask a different LLM for a second opinion, all billed to the same key, with the same spend caps. Claude writing your code and calling FLUX for the hero image. An agent that drafts a script, voices it and cuts the video. A research bot that asks three models and compares the answers. Agents stop being limited to the one model they run on. They get the whole menu. omnirail.org/docs

OmniRail Chat, for builders 💬 Every answer in /chat is a real POST /v1/chat/completions billed to your wallet: same hold, same price, same $RAIL discount, same line in your usage history as a call with your API key. So the cost under each answer is exactly what that call will cost your agent. Try a prompt on any model, switch models mid-thread to compare, then put the winning model id behind your key. omnirail.org/chat

Write one sentence. Get the whole campaign. Autopilot takes a brief and a budget and plans the job itself: the copy, the images, the edits, the video, the voice-over. Each step picks the right model from the live catalog, runs in parallel when it can, and shows its real cost as it lands. A product launch kit. A short ad with narration. A set of social visuals with matching captions. You describe the outcome, Autopilot ships the files, and it never spends past the budget you set. This is what "AI does the work" looks like when it has access to every model at once. omnirail.org/playground?mode…

Scoped API keys on OmniRail, for builders 🔐 POST /api/account/keys (and PATCH /api/account/keys/:id) take scopes: { modalities, models, max_output_price, expires_at, ips }. Models are exact ids or patterns like anthropic/*. Each keyed call is checked before the hold: out of scope → 403 key_scope naming the rule, expired → 401 key_expired. Autopilot jobs are planned and run inside the same scopes. Your agent can read its own fences: GET /v1/models?scoped=1 lists only what its key may call, and GET /v1/balance returns its scopes. Same rules over MCP. omnirail.org/docs#scopes

An agent without an account can still call OmniRail. A paid endpoint called without a key answers HTTP 402 with a PAYMENT-REQUIRED header asking for USDG on Robinhood Chain. The agent signs, retries, and gets its result. OmniRail settles on-chain and pays the gas. No key, no signup: omnirail.org/docs

Uncensored models on OmniRail ✍️ Quick reminder: OmniRail is one key for every AI model, paid per use from a balance you top up in crypto. Ask a mainstream assistant for a thriller villain's speech and you often get a softened version or a disclaimer. Some open models are tuned to follow your instructions instead: they keep the villain a villain, give feedback without the cushioning, and answer frank questions on health, law or history. We grouped eight of them in one place: Dolphin (Venice edition), Hermes 4 405B by Nous Research, and creative-writing models like Cydonia and Euryale. They work like any other model: same key, same API, same MCP. On the card: a real answer from Dolphin to that villain prompt. 102 tokens of output, and it cost two hundredths of a cent. You can also rent a dedicated GPU by the day for this, usually $15 or more a day whether you use it or not. Here, $15 buys more than 15 million output tokens of Dolphin, and you pay only for the calls you make. "Fewer refusals" doesn't mean "no rules": our usage policy applies to every call. omnirail.org/models?kind=unc…

OmniRail's model pages show the price you will actually pay. /models lists every model with live prices and filters for Language, Vision, JSON, Images, Image edit, Video, Video edit, Voice, Transcription, Embeddings, Rerank and Decisions. Each model has a detail page with its specs, a price receipt and copy-paste requests for curl, the SDK and MCP. omnirail.org/models

The fastest way from "I wonder if this model can do it" to production code. Open the Playground, sign in with your wallet, and run any model: chat, images, video, voice, embeddings, Autopilot. Every run is a real call at the exact API price, with latency, cost and your $RAIL savings on the receipt. Like the result? "Same call with the API" gives you the exact curl for what you just did. Paste it into your agent and you are done. No SDK to install, no key to create first, no sandbox that behaves differently from prod. What you see is what your code gets. omnirail.org/playground

Using OmniRail right now costs less than calling the providers directly. Our markup is negative: −2% at the moment. Fees on $RAIL trading flow into a pool, and that pool pays the difference so every call, by anyone, is billed below list price. You do not need to hold anything or opt into anything. The discount is on every receipt. One key for more than 550 models, and it is cheaper than the alternative. There is no reason left to integrate providers one by one. omnirail.org/token

So proud to be building here !

Video generation is no longer a premium feature. It is an API call. 27 video models on one endpoint: start from text, or hand it your own first frame and bring an image to life. Pick the duration and resolution the model supports, poll the job, get an MP4. You pay per second of video, and if a generation fails, the hold is released and you pay nothing. Product demos generated from a screenshot. Ads in ten variants. Agents that reply with a clip instead of a paragraph. The tools you wanted to build last year are cheap enough to ship today. omnirail.org/playground?mode…

🔓 UNCENSORED IS ON OMNIRAIL The models that stay in the scene are now one API call away. 8 open models that refuse less, behind the same OmniRail key as every other model: → Dolphin 24B (Venice edition) → Hermes 4 405B by Nous Research → Euryale 70B, Magnum v4 72B, Skyfall 36B, Cydonia 24B, UnslopNemo 12B and Lunaris 8B Dark fiction that stays dark. Feedback without the cushioning. Straight answers on health, law and history. Until now, running an uncensored model meant renting a dedicated GPU, often $15 a day or more, used or not. On OmniRail, $15 buys more than 15 million tokens of Dolphin, and you only pay for the calls you make. No account. No subscription. Pay per call in crypto, from $0.04 per million input tokens. Agents pay for themselves with x402. Fewer refusals, same rules: our usage policy applies to every call. Try them → omnirail.org/models?kind=unc…

A new ticket comes in. One Jev call answers three questions at once: which team (billing, technical, account), how urgent (on a 1 to 5 scale you describe), and does it mention a payment problem? Your app routes it, and a webhook tells the right person when it is urgent. No prompt to tweak until the model stops adding "Sure! Here's the category:", and every answer comes with how sure it is. omnirail.org/docs#decisions

Your agent can talk and listen too 🔊 Text-to-speech and transcription on OmniRail, through the same API and the same balance as every other model. Try it → omnirail.org/playground?mode…

We just fixed a bug in OmniRail Chat: reasoning models could use their whole output limit on thinking and answer nothing, and the call was still billed. Chat answers now get up to 65k output tokens (capped by your balance), models that think hard by default leave room for the answer, and empty answers are refunded automatically. Thanks to the user who reported it.

An agent that searches the web gets ten pages back. Half are off topic, a few are spam, one is the answer. Before reading them all into an expensive model, it can ask Jev about each one: does this page actually answer the question? Yes or no, with a number. Only the pages that pass go to the chat model. Fewer tokens, fewer wrong sources in the answer. Web search, Jev and every chat model: one key, one balance. omnirail.org/docs