Changelog
Every update that is live, newest first: what changed for you and your agent, and where to read more.
Chat
Trying several models used to mean one app, one account and one subscription each. Now /chat is a chat app for every text model: sign in with your wallet, ask, switch models in the middle of a conversation, stop, edit or regenerate an answer. Conversations follow your wallet on any browser, and each answer is a real chat completion billed to your prepaid balance at the API price, its cost shown under it, so you see exactly what the same call costs your agent.
Open Chat →Chat
Asking about a document meant copy-pasting it into the prompt. Now you drop PDFs, images, audio, video, text or code files into /chat, up to 10 per message, and ask about them. When the model you picked cannot read a file, the chat offers one that can in one click; reading them is billed like the same call on the API. Files are kept for 24 hours, then deleted (the conversation keeps its text).
Open Chat →Chat
Voice modes usually stick to one model. In /chat you can dictate a message, read any answer aloud, or switch to hands-free voice mode with any text model: the answer is spoken while it is still being written, and you can cut in. Each transcription and spoken sentence is billed at the API's voice prices.
Open Chat →Chat
With a model that can call tools, ask /chat for a picture, then ask for changes, in the same conversation; or tap Image or Video to pick the model and the format yourself. Videos, and any generation above the price you set, show their maximum cost and wait for your OK before anything is charged. Generated images and videos are kept for 24 hours: download the ones you want to keep.
Open Chat →Chat
Showing someone a good answer meant screenshots. Now Share turns a /chat conversation into a public, read-only link: anyone can read it without an account, costs are never shown, your wallet address only if you choose, and readers can continue the conversation in their own chat. You can update or delete the link at any time.
Open Chat →Keys
An agent's key used to open every model, with no end date. Now each key can carry its own rules: the modalities (trading included) and models it may call, a ceiling on the output price of text models, an expiry date and the IP ranges it may come from. A call outside those rules is refused before anything is held or charged.
Scoped keys in the docs →Agents
Long jobs used to mean polling: your agent kept asking whether its video was ready. Now it passes a webhook_url with a video, an Autopilot job or a trade, and OmniRail posts the signed result to it when the work ends, retrying deliveries that fail. The Webhooks tab of the dashboard rotates the signing secret, sends test events and redelivers.
Swaps
Your agent can now trade ten tokens launched on Pons, next to ETH, USDG and $RAIL. Each one is verified on-chain before it is quoted and leaves the list on its own if its contract or pool stops matching. They are volatile and listed for the depth of their pool, not endorsed; a key's token list allows them all with one entry, PONS_TOKENS.
Swaps
Agents had no safe way to swap tokens on their own. Now yours asks OmniRail for a quote between ETH, WETH, USDG and $RAIL, signs the transaction with its own wallet and pays its own gas; OmniRail never holds the funds or sees the key. On each key you set the tokens, the most per trade and per day, the widest slippage and the amount above which you approve each trade yourself. You can also swap by hand in the Playground.
How swaps work →Community
Updates now land on the OmniRail Telegram channel too, each with what it does and how to use it. OmniRail never messages anyone first: whoever contacts you in private in our name is not us.
Join the channel ↗Models
Choosing one model out of hundreds was your agent's job, on every call. Now it can send model: "auto" (or auto:cheap, auto:fast, auto:best) and OmniRail picks a model that can serve the request, moving to the next candidate if a provider fails before answering. The pick costs nothing extra, and every answer names the model that ran.
Payments
You could fund your balance with USDG or ETH; now $RAIL works too. It is credited in dollars at its on-chain price, the lower of the spot price and the 30-minute average of its Uniswap pool, and a deposit made while that price looks pushed around is held for support instead of being credited at a bad rate. You never need the token to use OmniRail.
Wallet deposits in the docs →Site
Getting help used to mean finding us on X. Now the chat bubble at the bottom right of every page reaches a human: sign in with your wallet to write, and the answer comes back in the same bubble. Errors in the dashboard and the Playground carry a Contact support button that attaches the error for you.
Errors and support →Site
The site now follows your system theme, day or night. Pick System, Light or Dark at the bottom of any page to change it.
Playground
Four more panels in the Playground. Generate videos, make a model speak or transcribe a recording, hand Autopilot a brief and watch it plan and deliver, and see what the $RAIL pool takes off your prices. Every run is a real call, billed like the API.
Open the Playground →Playground
Trying a model used to take code. The Playground lets you sign in with your wallet and talk to any chat model from the browser, then generate and edit images. Every run is a real call, billed at the API price, with its receipt.
Open the Playground →Keys
A key could already carry a total spend cap. Now it can also have a daily and a monthly one (UTC), editable from the dashboard at any time, so an agent stuck in a loop cannot spend a month of budget in an afternoon. A call that would cross a limit is refused before it runs.
Keys and limits in the docs →Dashboard
Where did the money go? The Usage tab of the dashboard shows what each key, each model and each day cost, from today back to 90 days, and exports every request of the period as CSV.
Open Usage →Models
A model only knows what it was trained on. Add web_search: true to a chat call, or web: true in the MCP chat tool, and any model reads live web results and cites its sources.
Models
After a search, your agent often holds dozens of passages and needs the few that matter. POST /v1/rerank, or the MCP rerank tool, scores each one against the question so it keeps only the best.
Agents
One provider going down used to cost you a step of the job. Now, when a provider fails during a step, Autopilot retries that step on another model of the same kind.
Autopilot in the docs →Models
Your agent can hand a PDF to any chat model, in the prompt or with files in MCP. pdf_engine picks how it is read: natively by the model, by free text extraction, or with OCR for scans.
Models
Memory, semantic search and RAG all start with vectors. Your agent now turns text, and images on multimodal models, into embeddings with POST /v1/embeddings or the MCP embed tool, with the same key and balance.
Agents
One brief in, a finished deliverable out. POST /v1/jobs, or the MCP build tool, plans the steps (copy, images, edits, a short video, a voice-over), picks a model for each, runs them in parallel within your budget and bills each step at its real price.
Models
Your agent can talk and listen: text-to-speech with POST /v1/audio/speech and transcription with POST /v1/audio/transcriptions, both OpenAI-compatible, plus the MCP tools speak and transcribe. Same key, same balance.
Payments
$RAIL launched on Pons. A share of its trading fees goes to a public on-chain pool, and while the pool has funds it pays part of every call, for everyone, with no token to hold. Every funding is listed with its transaction.
$RAIL and the pool →Payments
No account, no key, no card. A paid endpoint answers 402 with a price in USDG; the agent signs a payment with its own wallet, sends the call again and gets its answer, without paying gas.
Agents
Every provider used to mean another account, another key and another card. OmniRail puts text, image and video models behind one OpenAI-compatible API and one MCP server, paid in USDG or ETH on Robinhood Chain. Your agent sets itself up from one message: read omnirail.org/agents.md.
Quickstart →