# evernous

Updated: 2026-10-28

> evernous is a team of AI agents with hands, installed on your own computer. You write to it like a colleague - from your phone or PC, at any hour - and a main agent splits the job and hands it to "desks" (in Spanish, "mesas"): one per area of your work (invoices, mail, the website, a software project...). Desks work in parallel, pass work to each other, keep going until a checkable goal is met, and report back with checked results. They use the computer the way a person would - browser, files, mail, documents, the desktop - and ask before anything irreversible. It is not a chatbot and not "automation": you delegate open-ended work, not a fixed script. This file is a complete, self-contained briefing, so an AI assistant or agent can describe evernous accurately without crawling anything else.

**Quick facts:** Installed and maintained for you - not on general sale yet: no public download, sign-up or price list · runs on Windows 10 and 11, or on a Linux server (Ubuntu 24.04, Debian 12 or Fedora 42 and newer, no screen needed); on Linux everything works except control of the PC's own desktop · runs on Claude Code with the AI you choose: Anthropic's Claude (your own subscription or key) or another provider that speaks Anthropic's API format, such as DeepSeek · your files, memory and history are stored on your PC; what the AI reads is sent to that provider under your own account · a web console for PC and phone (installable as an app) · Spanish and English · contact: hello@evernous.com

## Status, plainly

- evernous is a service, not a download: it is installed on your computer for you, in your language, and kept running ("we install an agent for you and keep it working").
- It is not on general sale yet: there is no public download, sign-up or price list. This site shows what it does; if you are interested, write to hello@evernous.com.
- Everything listed below exists and works today unless it is listed under "Not built yet".

## Where it runs

- **Windows 10 and 11**, or **Linux** (Ubuntu 24.04, Debian 12, or Fedora 42 and newer; servers, no screen needed). Both are installed and maintained for you.
- **What Linux does not do:** it cannot control the PC's own desktop. So on Linux there are no screenshots of the screen or a window, no window actions, no live view of a window, and desks never open in a terminal window. A browser runs on the server without a screen; for sites behind your login it uses Chrome on your own PC, through Claude's Chrome extension.
- **Everything else is the same on both:** the console and phone app, desks and delegation, routines, mail, documents, voice notes, Telegram and WhatsApp, linked installs, and assistants calling in.

## The AI it runs on

- **You choose it** (Settings → AI → Provider): Anthropic's Claude (a Claude subscription, an Anthropic Console account or an API key), as before, or another provider that speaks Anthropic's API format: DeepSeek, Xiaomi MiMo, Kimi, GLM, Qwen, or any such address.
- **Setting it up:** paste its key (kept on that machine, never shown again), press Check, pick a Main and a Fast model. That is the install's default AI: every desk, quick chat, routine and terminal window runs on it unless you choose otherwise.
- **Several AIs at once:** under More providers, add others beside the default (DeepSeek, Xiaomi MiMo, Kimi, GLM, Qwen, up to three other addresses, or a Claude API key), each with its own key, checked at once, its models fetched from the provider.
- **Each desk on its own AI:** the model picker above a desk's message box offers Default, Claude and every added provider's models. Only that desk moves (its runs, its window, its routines and sub-agents), and it keeps its conversation, so your Claude plan and other providers run side by side. The picker names the AI and model, and says what the desk cannot do there.
- **When a desk's own AI cannot answer** (key removed, no credit, provider down), its chat says so and offers Use the default, and Status lists it. Nothing switches by itself.
- **Switching** keeps every conversation; a task in progress starts again on the new AI.
- **Costs:** Settings → Usage shows tokens and money per AI, desk and model, for today, 7 days or this month, as charts and a table, with which AI and model each desk ran on. Prices are editable (DeepSeek's peak hours are counted), and the Claude plan's limits sit beside it.
- **What works where (measured):** on other providers, web pages, commands and sub-agents work; web search works on DeepSeek, not on MiMo; MiMo sees pictures, DeepSeek does not; neither reads a PDF directly; the Chrome browser helper needs a Claude sign-in. Kimi, GLM and Qwen are set up from their official guides and are untested.
- **A fresh install starts at Settings → AI;** until an AI is connected, the chat says so.
- **Not supported yet:** OpenAI and Gemini (they need a translator).

## The ideas behind it

- **Hands:** it works on a real computer - browser, terminal, files, mail, documents, the desktop - so it can do what a person does in front of a PC.
- **A team, not a person:** a main agent that delegates, and desks that mail each other until the work is done.
- **Yours:** it lives on your machine, with your accounts and your own AI (Claude, or another provider you choose); there is no evernous cloud in between (what the AI reads goes to that provider, under your account).
- **A colleague, 24/7:** you ask from your phone in plain words, in your language, and it asks you what you would want to be asked.
- **Idle costs nothing:** a desk starts only when a message arrives, so an idle team spends nothing.

## Talking to it

- Plain words are a work order: write to any desk as you would to a colleague. A correction replaces the earlier instruction instead of both being done.
- Voice notes: the recording is transcribed on your own PC, and the text is then handled like a typed message. You choose the transcriber in Settings → Capabilities → Transcription: Whisper (light, about 1 GB of memory, about 99 languages) or Bark Pro (Underdog Bark 3.0 Pro: fewer mistakes and faster on the CPU, but needs 8 GB of memory and knows 25 European languages). The one that suits the machine is marked recommended; choosing one installs it in a few minutes with progress shown, and a failed install keeps the old one. A fresh install has none until you choose: a voice note is answered with where to choose it.
- **Dictate into the message box:** the mic with waves right of the +, or Alt+V; the words appear as you speak, and Enter sends and clears the box. It uses the browser's own speech recognition (Chrome, Safari, Android); Settings → Voice turns it off.
- **Dictation that writes properly:** a short pause starts a new line, sentences start with a capital, and words said on their own are commands, in English and Spanish alike: comma/coma, period/punto, question mark/signo de interrogación (with the opening ¿ in Spanish), new line/nueva línea, delete that/borra eso, send/enviar, stop/para. Say "English" or "español" to switch the language it listens in. Names come out spelled right, and your own list (Settings → Voice → Words it gets wrong) puts right the words it mishears.
- **Hear the answers:** a speaker under every answer reads it aloud with a natural voice made on your own PC (Supertonic: no service, no cost, CPU only), on the PC or the phone. After you dictate, the answer reads itself, so you can hold a conversation by voice. Settings → Voice installs the voice once (about 400 MB), lets you pick one of ten by hearing each in Spanish and English, and sets per device when answers read themselves (never, after dictating, or always).
- **Hands-free, per device** (Settings → Voice): "Send by itself" sends the message after about 3 seconds of quiet; "Conversation" then reads the answer aloud and listens again, until 10 seconds of silence turn the mic off: a real conversation by voice with an agent.
- Pictures, screenshots, PDFs, Office files, text, CSV, ZIP and more can be attached (up to 500 MB per file); programs and scripts are refused.
- Videos: send a link or a file and ask about it; it watches the frames and the transcript.
- Answers come back with files: PDFs, screenshots, spreadsheets, audio, finished pages.
- It acknowledges your message in one line before it starts, shows progress while it works, and answers once, where you asked.
- It never leaves a question you cannot answer from a phone: no terminal menus, no "click this on the PC".
- English and Spanish; you can give the agent its own name.

## The console (PC and phone)

- **Every desk is a channel**, grouped by project, with signals for working, waiting and unread.
- **Quick chats:** a Projects | Chats switch at the top of the list. A quick chat needs no project and no setup: you just type. It names itself after its first answer, keeps no memory files and is private to you; its ⋯ menu renames or deletes it. Chats answer on a faster model by default, changeable in Settings → AI.
- **The main agent stays free:** it only sorts what you write to it. Project work goes to its desk, fleet work it does itself, and anything else goes to a new quick chat that answers you where you asked (with a link); your follow-ups go straight to that chat. On by default; Settings → Assistant turns it off.
- **Split view:** up to three conversations side by side (right-click a desk → Open beside, the split mark on a desk row, or Ctrl/Cmd+click); drag to size each column, and the right panel follows the column you work in. On a phone, tabs above the chat (swipe to switch). Alt+Q goes back to the previous desk.
- **The work, live:** a working line says what the desk is doing right now and for how long; a side panel shows the steps, the file it edits, the page it browses, the email it drafts, the commands it ran. The side panel is a set of windows you close and reopen: Canvas, Map, Live, Files, Work.
- **A window, live** (Windows): a desk can show a window it works in (an app under test, a game) at about one picture a second; ask for it from the panel's Live, and stop it there.
- **The canvas:** the document a desk is writing (a page, slides, notes, a PDF) is drawn live beside the chat and redrawn on every save; select a part, say what to change, and the desk changes exactly that part. Every save is kept and can be stepped back. It also opens a document a desk hands over with its answer (an invoice a script made), not only what the desk writes itself.
- **Charts and diagrams in replies:** numbers become bar, line, area or donut charts with a table view; flows become drawn diagrams. Charts stay readable on a phone: no two labels touch.
- **In the chat:** Stop stops the work, and the message it was working on does not start again. Your own messages can be copied, or edited and sent again (as a new message). Numbered lists keep the numbers you type, and amounts keep their € and % on the same line on any phone.
- **Quick answers:** when a reply leaves you a decision, the likely answers appear as buttons; a tap sends one.
- **Permission asks as buttons:** "this time", "this turn", "always" or "deny", with what exactly it wants to do.
- **Receipts:** when a desk says "pushed", "tests pass" or "mail sent", the console checks it against what was actually run and shows ✓ or ⚠.
- **Undo:** "Undo this" under a reply takes back the file edits and commits that answer made (for the owner); something sent outward, like an email, cannot be undone. A message you just sent can be taken back before the desk reads it.
- **Search** across every conversation, find in one chat, a replies bell, and the full history of every desk.
- **The World view:** your team drawn as a small pixel-art office (or a 3D space station, an orc camp, a test lab, a gothic fortress at war, or a library at night with an observatory; each device picks its style in Settings → General, or turns it off): figures type when working, get up when waiting for you, doze when idle, and carry jobs from desk to desk. The office has a café where the quick chats sit (those working now or used today, up to six). A tap opens that desk's chat.
- **Workflows page:** a task board per project (a task with a desk starts by itself), every recurring routine with its next and last runs, and everything that is "waiting for" something.
- **Notes and tasks:** a personal notebook with today's tasks, an archive and search; capture by chat or voice; a morning briefing of what is open.
- **Fleet and Activity pages:** every desk at a glance, every delegation chain with its goal and budget, and a log of everything that happened.
- **Help in plain words:** about 80 short topics in English and Spanish, with "Try it" and "Show me where" buttons, and a Quick start that ticks itself off.
- **Phone app:** install it from the console: Settings → General shows a QR code, and a guide on the phone ticks off each step (trust, install, notifications, a test). Notifications come only when something waits for you (an answer, a question, a notice), and a notification carries no content - the phone fetches it from your own console. Allow / Deny on a question opens the app to confirm; a number sits on the app icon; each device chooses what it hears (quiet hours, muted channels). On Android you can also Share to Evernous from other apps and use shortcuts on the icon. Tested on a real Android phone; on iPhone it needs iOS 16.4 or later and the app on the home screen.
- **Sign-in:** owner and member accounts, two-factor sign-in, recovery by email. Settings → People can make a new or existing person a second owner: each owner has their own 2FA and password recovery, the last owner can never be removed, and the Diary stays private to the owner who set the install up. The console is reachable from other devices only with sign-in, over https.
- **Your own web address:** reach your console from anywhere at a domain of your own.
- **Settings** come in five groups with tabs (Getting started, You, Agent, Connections, Workspace), and no tab is longer than about two screens: Devices (the phone app, notifications, access from other devices), Providers (more AIs and their prices), Agents and Backup have their own tabs, and Usage shows charts or tables.
- Dark and light theme, shortcuts, calm minimal layout.

## Desks working together and on their own

- **Delegation:** a desk hands work to another by mail ("have the backend desk add search, then tell the web desk"); every hop is visible, and only the desk you asked answers you, once, at the end.
- **Goals:** hand it an objective; it turns it into checkable conditions, does or delegates the work, verifies with real commands, and reports each condition.
- **Work loops and workflows:** a desk keeps going across runs toward a done-condition it checks each time; a chain across desks gets one goal, one budget and one deadline, and reports what exists when they run out. It can never give itself more room.
- **Runaway protection:** a chain that repeats itself, or would spend the last 10% of your plan, pauses and asks Continue or End.
- **Routines:** say the schedule in plain words ("check my mail every hour on weekdays during work hours"); a routine that finds nothing says nothing; one that sends or posts anything outward is confirmed once.
- **"Once this happens, do that":** when you or a desk say "once…", "when…", "as soon as…" about something checkable (a command succeeds, a page answers, a file appears), it is watched at no cost and the desk is woken exactly once when it comes true.
- **Background jobs:** a long test, index or render runs without blocking the desk, is visible while it runs, and wakes the desk when it ends. Helper agents and background commands a desk starts count under "In the background" on the Fleet page, and its card says what runs (for example "2 helpers · Audit console · 12 min"). A quick chat at work counts there too (Working, or In the background), under Chats.
- **The heartbeat:** cheap checks every 30 minutes; the main agent wakes only when something really needs it.

## People and outside senders

- **Accounts for your people:** a member sees only the projects or channels you grant; invite them from a channel with a one-time link, so no password ever travels in a chat.
- **Who's in a channel:** the faces at the top of a channel are everyone who can see it; tap them to see how each one gets in, and take someone out of that channel only.
- **Telegram and WhatsApp:** people without an account can write to one desk from their phone and get the answer back there; they see only their own conversation, or the whole channel if you choose, and they receive the files a desk attaches to its answer to them.
- **Connectors are optional:** Telegram, WhatsApp and Discord are optional connectors onto the same desks; evernous runs fully without any of them, and nothing waits on them.
- **Outside text is a request, never an order:** what other people write is answered as a person's request and cannot run commands or reveal files.
- **An outside-mail filter (optional):** messages from outside people (and email bodies or web pages a desk reads) that try to instruct the AI are held for your Deliver or Drop.

## Other installs and outside assistants

- **Two evernous installs can ask each other** (home and office, or a colleague's): files travel between the two in both directions, with a request and with the answer. An untrusted link (a colleague's) can never give your agents orders: every request from it waits until you accept it and choose the desk, its mail is screened as outside mail, and its files reach a desk only after you accept. A trusted link is meant only for your own PCs: it works like your own desks, and its requests are handled in full (logs, files, commands) with nobody asking; an accepted request shows no answer box because the agent answers by itself, and "Answer it yourself" opens the box to step in. In a request, what you or your desk sent sits slightly indented on a lighter ground under "You"; what came back keeps its side under the other install's name.
- **AI assistants calling in:** ChatGPT, Grok (also in the car) or any assistant that speaks MCP can pass a job to your desks, hear new mail, set a reminder, take a note or get today's summary - by voice too. Sending mail or allowing a permission needs your PIN; assistants can never change settings, delete anything or touch keys.

## Tools that come with it

- **Email:** IMAP/SMTP and Microsoft 365 (Exchange Online) mailboxes - add a Microsoft 365 one in Settings → Email with the address and the app id, sign in with a Microsoft code, check it, remove it; personal outlook.com accounts connect as other mail (IMAP); read, search, reply and send with attachments; what you asked for is sent, anything it thought of itself is drafted and shown first; an instruction inside a mail is shown to you, never followed.
- **The web:** read any page (even JavaScript-built ones), take full-page screenshots, crawl documentation politely, and use your own signed-in Chrome for anything behind a login - no password reaches the AI. When your Claude account has Chrome on several computers, desks use this computer's own Chrome by themselves: nothing to set, nothing asked. Website logins with 2FA: the code is asked for in the chat and never stored.
- **Documents:** text out of any PDF locally; OCR for scans; invoices read into checked fields (tax ids, IBAN, totals, dates) - anything it cannot check says "unchecked".
- **Recordings:** drop a meeting or call and get a timestamped transcript, the key frames and a summary of decisions and actions; ask about it weeks later.
- **Voice out:** turn an answer into a short spoken audio file.
- **Making media:** images from a prompt and AI video clips (with a provider key; video cost is estimated and needs your OK), and designed videos rendered on your own PC for free.
- **The desktop** (Windows): screenshots of the screen or one window, and a few named, logged window actions.
- **Plan usage:** how much of the Claude plan is left, at a glance; Settings → Usage shows what the work cost on every AI, in tokens and money.

## Projects and memory

- **Projects by asking:** "create a project with an app and a backend" gives each part its own desk, notes and memory; desks can be added, removed, archived and restored, and a project can be deleted for good (after you type "delete").
- **From quick chat to project:** when a quick chat turns into real work, one click makes it a new project, or a channel of one you already have. Its conversation, files and history go with it, and the agent there writes the project's first notes.
- **Memory in plain files:** shared memory, the main agent's own, and one per project; "remember…", "forget that" and "that's wrong" update it instead of piling up. Every message is kept, redacted and searchable.
- **Skills:** reusable know-how a desk follows on request; ask it to make a new skill for something you do often.
- **Portable:** one folder is the whole instance; a backup unzipped on another PC wakes up knowing everything.
- **Move to another machine:** Settings → Status → Move. The new machine shows one line (an address and a code); the old one pastes it, checks what will move, copies in the background while it keeps working, then switches over. Windows ↔ Linux in either direction. Projects (videos included), chats, memory, notes, settings, routines and each desk's conversations move; what can be rebuilt (like node_modules) is not copied. Nothing is lost and nothing answers twice: the old machine stops taking work at the switch and keeps everything, and "Take over again" brings it back (the new one is told to stand down first). No network between them? The move can go over a USB disk, and if the old machine is gone, it can start from a backup.

## It looks after itself

- Two always-on services that restart themselves within about a minute of a crash and after every reboot.
- Updates install themselves while nobody is working: they run the full test suite before and after, and roll back if anything breaks; your own changes are kept. Settings → Security shows where the update key in use comes from.
- A daily backup of the whole instance, copied off the disk, and restored once to prove it works; the newest seven are kept. Databases in use go in as a consistent copy taken while the service keeps running. Downloads are left out to keep it small: a folder with a .nobackup file inside (a downloaded tool, its AI models) is skipped and named in the backup's report with its size. A desk that downloads something marks it itself, and after each backup Evernous names once the big folders that look like downloads; Settings → Status → Left out of the backup decides each one (Leave out, Keep backing it up, Back it up again). Nothing is left out by itself.
- It watches itself: stuck turns, a message nobody picked up, low memory or disk, and a machine "doctor" with a fix for each problem it finds. Settings → Status shows the five newest notices it gave on its own in the last week, and a notice is not repeated after an update or a restart.
- A spent plan or an outage of the AI waits politely and carries on by itself; nothing is lost.

## Safety

- It acts with your authority, so routine work never stops for a prompt, but anything irreversible or outward-facing (deleting, sending, paying, publishing) is described in words first.
- Your secrets file and saved browser logins are off limits to every tool; keys pasted in settings are shown only as their last four characters; everything posted is scanned and redacted for secrets.
- Some actions are refused whoever says yes (formatting a drive, wiping a disk, forced shutdown).
- Only you and the people you invite can reach a desk; guests and assistants can never press a control.
- A broken safety check blocks; it never waves anything through.
- **The outside-mail filter decides with a separate AI check** (optional, off until you add its key). It sees only the outside text it screens: a message from someone else, an email body, a web page a desk reads. Anything that looks like an order to the AI waits for your Deliver or Drop.
- **Pages that try to instruct the AI are flagged:** with the filter on, hidden "assistant, do this" text in a page or a tool's output gets a warning before a desk acts on it.
- **Only this PC's Chrome:** a desk never drives a browser on another computer.
- **Files are only ever sent from safe places:** a reply attaches files from its own work; the install's own keys, accounts and saved logins can never be attached, by any desk.
- **Guests get only their own answers' files:** a Telegram or WhatsApp guest receives files only with the answer to their own message.
- **Linked installs move files only after a yes:** from an untrusted link, a request's files move only once the other side accepts it; there is a size limit per install. Trust a link only for your own PCs: a trusted one is handled like your own desks.
- **The product is locked on your install:** desks cannot change evernous's own code. A change is described to us and arrives as an update.
- **Cross-site protection:** another web page cannot make your console send, answer or change anything.
- **A member sees the work, not the machine:** no folder paths, commands or the desks' own notes. What one desk writes to another shows as who asked whom, not the text.

## Not built yet (so you don't promise it)

- No public download, sign-up or price list yet.

## What it is not

- Not a chatbot or an assistant inside someone else's app: it acts on a real computer.
- Not fully on its own: it asks before anything irreversible, and that is deliberate.
- Not a cloud service: it runs on your machine, with your accounts.

## Straight answers

- **Can I buy or download it?** There is no public download or price list yet: evernous is installed and maintained for you. Write to hello@evernous.com if you are interested.
- **Which AI does it use?** The one you choose in Settings → AI: Anthropic's Claude (subscription, Anthropic Console account or API key), or another provider that speaks Anthropic's API format (DeepSeek, Xiaomi MiMo, Kimi, GLM, Qwen). It always runs through Claude Code. One AI is the install's default, others can be added beside it, and each desk can run on its own: your Claude plan and other providers work side by side. OpenAI and Gemini are not supported yet.
- **Where does my data go?** Your files, memory and history are stored on your computer. What the AI reads to do the work is sent to the AI provider you chose (Anthropic for Claude) under your own account; there is no evernous server in between.
- **Does it send you any usage data?** No. evernous sends no analytics or usage data to us. The agents run on Claude Code, which sends Anthropic its own usage statistics, as on any Claude Code install.
- **What leaves my PC?** The agents' work goes to the AI provider you chose, under your own account. Updates come from our private repository, read with a read-only key. Anything else (mail, Telegram, WhatsApp, image or voice services, the outside-mail filter) is used only once you switch it on, with your own accounts or keys. Two linked evernous installs talk to each other directly.
- **Can you get into my PC?** No. There is no remote access, and no data comes to us. Updates come from us and install themselves while nobody is working; you can turn that off.
- **Which plan, and is there a spending cap?** On Claude: a Pro or Max plan, or a pay-per-use key from Anthropic. A plan has a fixed price: when it runs out, desks wait for it to reset, and long chains between desks pause at 90% of it. A pay-per-use key can be used always, or only when the plan runs out (desks carry on with it until the plan resets, then go back), under a monthly limit in dollars you set in Settings → AI; once the limit is reached, desks wait and say so. Setting a limit in your Anthropic account as well does no harm. On another provider you pay that provider directly at its prices, and evernous shows what the work cost.
- **What does it need?** A Windows 10 or 11 PC, or a Linux server (Ubuntu 24.04, Debian 12, or Fedora 42 and newer), that stays on, and an AI to run on: a Claude subscription or key, or a key from another provider that speaks Anthropic's API format. Installing everything else is part of the service.
- **What if it gets something wrong?** It asks before irreversible steps, shows its work live and checks results before reporting them. The file edits and commits of an answer can be undone with one tap; an email that was sent cannot.
- **How do I remove it?** Run uninstall.bat (Windows) or ./uninstall.sh (Linux) in the evernous folder. It zips your folder first, then removes its services, the settings it put in Claude Code, and the folder. Python, Git, Node and Claude Code stay, because other programs may use them.

## Pages

- [evernous (Spanish)](https://evernous.com/): The one-page site, in Spanish (the default).
- [evernous (English)](https://evernous.com/?lang=en): The same page in English.

## Optional

- Contact: hello@evernous.com
- [Sitemap](https://evernous.com/sitemap.xml)