Changelog
Every release, newest first.
Connectors, step by step
Connecting a service is now a short sequence of steps, and a service is only added once every step is done.
- ImprovedConnectors shows the steps each service needs: sign in, paste a key, or nothing; then choose what Untold may do; then it is added.
- ImprovedA service is not added until you have chosen at least one thing it may do. Until then it is not counted, not offered under the message box and not used in answers.
- ImprovedConnections started earlier but never finished now show as “Finish connecting”, with Continue and Cancel.
- ImprovedLooking at a service in Connectors no longer creates anything.
askuntold.ai
Untold moves to its own address, past agent runs can be reopened, GitHub sign-in works from the agent, and errors say what went wrong.
- NewUntold is now at askuntold.ai. Links to askuntold.com forward there, with the page you asked for.
- NewSign in to GitHub from the connect dialog with a one-time code, without leaving the agent.
- NewPast agent runs are listed beside the agent. Reopening one brings back its answer, its steps and the services it contacted, and a follow-up continues the same run.
- ImprovedSheets and dialogs in the Studio and the Feed keep the keyboard inside them, close with Escape, and return you to where you were.
- ImprovedWhen a model cannot answer, the reason now reaches your screen instead of a generic error.
- ImprovedThe home page animation loads on more browsers, and shows a still frame where video cannot play.
- FixedFor a few hours after the move to the new address, answers failed to arrive. They are back.
- FixedA message typed while a saved conversation was still opening could be saved in its place. Sending now waits until the conversation is on screen.
Connectors and skills
The agent can work with the tools you already use, on your terms, and learn how you like a kind of task done.
- NewConnect your own tools, such as GitHub, Notion, Linear or any MCP server, and tick exactly which of their actions the agent may use.
- NewCredentials for tools you connect stay in your browser and are sent only with the request that needs them.
- NewSkills: write instructions for a kind of task. The agent loads one when the work matches, and each answer shows which skill shaped it.
- ImprovedEvery answer names the services it contacted.
- ImprovedThe account page is organised into sections: connected tools, skills, plan, identity, wallet, invites and your data.
Receipts wherever an answer appears
Checking an answer no longer needs anything from us: the proof sits under the answer, and more of the agent's work carries one.
- NewA Verify button under every sealed answer checks its receipt in your browser, without sending a request.
- NewAgent answers built from several steps carry one receipt for the whole run, with a link for each model call.
- NewAttach a wallet address to your identity, so tokens you earn have somewhere to go. It is not a login and nothing reads its balance.
- NewInvite links with a copy button. Referral rewards are paid only when the person you invited pays for something.
- ImprovedAgent answers that needed no tool carry the same receipt as chat.
- ImprovedMore models run sealed: GPT OSS 120B, GPT OSS 20B, Muse Glimmer 30B, Nemotron 3.5 Lightning and Qwen 2.5 7B.
- ImprovedAnswers stream as they are written, with tools turned on.
- ImprovedFormulas render as mathematics, money amounts stay as written, and wide tables and formulas scroll inside their own box.
- ImprovedThe attestation indicator tells apart live, not checked yet, stale and failed.
- ImprovedEvery page is served with a strict security policy, and code the agent runs executes in a sandbox in your browser.
- FixedStop now works on the first message of a session.
- FixedAn answer that never arrives says so and offers a retry, instead of waiting forever.
Encryption on your device, and accounts
Your prompt can be encrypted before it leaves your browser, and your identity can follow you to another device.
- NewEnd-to-end encryption on sealed models: your prompt is encrypted in your browser and relayed unopened. On this path we cannot moderate what you send, and we do not store the conversation.
- NewSign in with an email and password to keep your identity across devices. Signing out revokes the session.
- NewPlans: Pro, Pro+ and Max, with unused credits rolling forward. Sealed models, receipts and encryption are on every plan, Free included.
- NewTalk to characters, not only browse them.
- ImprovedAnswers render their formatting: bold, lists, quotations, tables and code.
- ImprovedOnboarding offers only the choices the product actually has.
- ImprovedModels that cannot answer are left out of the menu.
- FixedFor a period, first-time visitors could not use Untold at all: sending a message failed without an identity. Sending now creates one.
First release: sealed, fail-closed, receipted
Private answers from models running in attested hardware enclaves, with a receipt you can check yourself.
- NewSealed models run inside attested hardware enclaves. The sealed tier is fail-closed: a request that cannot be attested is refused, never quietly downgraded.
- NewEvery answer carries a signed receipt you can verify without trusting us. It commits to your prompt with a salt only you hold.
- NewAn OpenAI-compatible API with keys, rate limits and a receipt on every response.
- NewModel traits: ask for default or most_uncensored instead of a model id, and your integration keeps working when the model behind it changes.
- NewA limits page that states what our privacy does not cover, computed from the running system.
- NewCredits are kept in an append-only ledger, and a balance can never go negative.
- NewPaid plans unlock topic freedom. Privacy is on every plan.