Everything Nova can do

The brain runs all the time and thinks with Claude models. A small Nova backend runs the dashboard window and passes requests to the brain. n8n runs on the same laptop.

At the start of every conversation the brain reads a persona file. That file holds Nova's character, facts about Aman taken from his own website, and how Nova should introduce itself. Nova is told that everything it says about itself must be true.

Before a dashboard request (typed or spoken) reaches the brain, a tiny judge model sorts it in about a third of a second, with simple word rules as a backup. Only that one run changes model. The chat shows a line such as "Opus 5.5 takes this one (building and code)." On Telegram, Aman switches models himself with /model.

Hermes can start helper agents by itself. On 14 October Aman set their model to Opus 5.5. Nova decides for itself when helpers are worth it, so on small jobs it often does the work alone.

After each conversation the brain reviews it and updates Nova's notes about Aman. Nova and a built-in skill keeper edit its skills. Every change is recorded, and the old version is kept. Each change is one line on the Chat page, and every evening at 21:00 Aman gets one short Telegram message listing what Nova changed in itself that day (none on a day without a change). The lines name a skill or say "its notes about you", never what a note says.

The backend keeps a short register of running background work and the open task list. A small add-on in the brain gives that register to Nova at the start of every new conversation.

Nova's services start when the laptop boots and restart by themselves if they crash. A health-check tool checks connections, keys, disk and memory. Planned restarts happen only when Nova is idle, so no request is cut off.

Since 18 October the brain starts after n8n and waits up to a minute for it to be ready. If n8n is missing, the brain still starts.

A small backend on the laptop collects live data and sends it to a window that opens like an app.

The background grid fades in, the centre core lights up and its rings draw outwards, the bars draw their corners, the panels slide in from both sides, the gauges sweep up from zero, and the state chip says ONLINE. Any key skips it. Reopening within 15 minutes is quicker, and with reduced motion it is a short fade.

It is drawn live, not played from a video. Idle is a slow blue breath. Listening is cyan and reacts to the voice. Thinking is indigo to violet with a sweeping light. Speaking is a bright blue pulse that follows the voice. Working shows green-to-cyan beams. Waiting for approval is amber and still. Error is red, and offline is grey. Beside the orb, the left side shows the voice (microphone, sound level) and the right side shows the work (current tool, step, time, model and a Stop button).

Three columns, plus a bottom strip for recent messages and the event log. Panel sizes adjust to the screen, and on smaller windows the columns fold into tabs. Since 26 October this page is called the Dashboard: it opens on a calm overview (below), and All panels brings these three columns back.

The backend reads the machine every second and every five seconds. The panel shows processor load with a 60-second chart and one meter per core, plus graphics, temperature, rings for memory, swap, disk and battery, Wi-Fi signal, and download and upload speed. Amber and red appear only when a reading is really too high or too low.

The backend really tests each service and shows one row each with a green dot, plus a summary such as "7/7 OK". A service that is down turns red, and the bottom bar shows a red count.

The NOVA wordmark is in the brand gradient, with a state pill such as THINKING or NEEDS APPROVAL. A MIC badge shows whether Nova can hear, and a click switches the microphone. ECO, PRESENT and DEMO badges show the active modes.

Click a page name, or press its key: Ctrl+1 Home, Ctrl+2 Chat, Ctrl+3 Automations, Ctrl+4 Usage, Ctrl+5 System, Ctrl+6 Projects and Ctrl+8 Social (Ctrl+7 is kept for the coming Voice page). Calendar, Settings and Help open from their tabs, and Help also with the ? key. Home shows a count of approvals and alerts, Chat breathes while Nova works, Automations counts failed runs, Usage warns about its figures, and System counts services that are down. The page never switches by itself. The small mark on a Home panel opens that panel's full page.

The window talks to the local backend, which starts each request on the brain and streams the steps back. The text box takes several lines: Enter sends, Shift+Enter starts a new line, and half-typed text follows Aman between pages. Long answers fold with "Show all". A web address in an answer becomes a small chip with the site's real name, and a click shows the whole address. It is never a live link and nothing is fetched for it, for safety. An "Open in browser" button waits for Aman's yes (one switch).

A paperclip, a paste (Ctrl+V) or drag and drop attaches up to four pictures. They stay on the laptop, and Nova opens them with its own tools. The chat shows "[1 screenshot attached]".

Nova ends its answer with the plan in a fixed format, and the backend keeps it. The Plan panel shows each step with its status in words (to do, doing, done), who owns it and a note. Nova can tick steps off in the same plan later.

Each action (terminal, browser, search and so on) is a row, and helper agents' work is nested under it. A finished task shows DONE or FAILED, how long it took and roughly what it cost. Work started from Telegram, WhatsApp or a schedule shows too, with its channel.

The mark breathes when idle, sweeps when thinking, rings when acting, and turns a still amber when it waits for Aman. Until the answer's first words arrive, a working line under the chat shows the current action, the time and the number of steps. It says HELD while waiting for approval.

The card shows on Home and on Chat. On any other page a line under the top bar says "Nova needs your approval" with a "Show it" button, and the page never jumps by itself. Y works only when pressed on its own, after a quiet second, so typing a sentence never approves anything.

Each alert is a card with an icon and a word, never colour alone, and an acknowledge button. Key A jumps to them, and Home's tab shows how many are open.

The tones are made inside the page (no sound files, no cost), each under a quarter of a second: heard, done, failed, approval, warning, critical, wake and boot. At night only approvals and critical alerts sound, unless Aman is at the laptop. Key S mutes them all.

Press P, or switch it on in Help. Conversation text (Nova's answers included), messages, commands, job names, costs, file names and calendar titles turn into dots, while the layout, counts and animations keep moving. A PRESENT badge shows in the top bar, and only the important sounds stay on.

A single letter acts only when pressed on its own, after a one-second wait (the bottom bar names the key meanwhile). A sentence typed into the window by mistake does nothing, and the bar says "Press / to type to Nova".

Larger vitals charts, every service with its own sentence, facts about the laptop (uptime, fans, disk temperature, firewall on), the five busiest programs, and the whole event log with an "All / Warnings" filter.

The page reads the project list and the milestone list (M0 to M26). Milestones can be filtered as Active, Blocked, To do or Done, and a project that has been quiet for seven days is marked STALE.

The dashboard reads Nova's scheduler. The next job is highlighted with a ticking countdown. Paused jobs are grey, a failed last run is red, and a missed time is marked "late". By design, nothing can be paused or run from the window.

It holds an owner's guide (still a first draft), the key table, switches (presentation, sound, reduced motion, eco, microphone, developer tools), a "Right now" status and "What you can say". The ? key opens and closes it.

Full mode runs the orb at 60 frames a second. Eco mode (on battery, when hot, when frames are slow, or with key E) drops to 30 or 10 and removes the glow. In the background it drops to 15, when hidden it stops, and after 10 idle minutes the decorations dim. Key R gives reduced motion.

Below a certain window size the full layout switches to a slim widget that keeps the essentials in view beside other work.

An amber or red "LINK STALLED" or "LINK DOWN" banner shows the time of the last data and the next retry. Panels say "Not connected yet" or "Not readable right now" in the backend's own words.

Open the window's address with ?demo=1 (or demo=today, demo=empty). Extra switches show a chosen orb state (for example state=thinking), an approval card, a long answer, an error or a lost link. No backend is needed.

One theme sets the colours: blue for the brand, green for alive, indigo and violet for working, cyan for listening, and amber and red only for warnings. It uses the Geist fonts, light beams travelling around active cards, the orbit ring, and a shimmer on changing numbers, never more than one pulse a second. Every state also has a word, never colour alone.

One theme tool applies every setting from the same colour file as the dashboard, checks that each one took, and can undo all of it. The fonts stay the system's own, which was Aman's choice.

It is a Claude Code plugin installed for every account on the laptop. It uses the dashboard's own figures, so both screens always agree. The /nova look command switches between three looks.

The lines sit in the Task box under Background work, headed "NOVA CHANGED ITSELF" with how many changes the last 24 hours brought. The backend reads the brain's change record: which skill changed, what kind of change, and when. For the note files it reads only their names and times, never their words. Quick repeats on the same skill or notes fold into one line with a count, and presentation mode hides the skill names. Every evening at 21:00 the same lines also reach Aman as one short Telegram message, the day's digest (none on a day without a change).

The dashboard could already draw these. Since 19 October Nova's writing rules for the dashboard ask for the richer form, and a small tool can switch them back. Telegram and WhatsApp keep their own rules.

Since 20 October the backend asks n8n itself, every few seconds, for its list of automations and runs (never what a run held). The 7 days come from the history the laptop writes every night, because n8n itself forgets after 7 days. Nova's own builder wrote it; Claude reviewed it and merged it.

Since 20 October 10:46 the backend adds up its own cost records per Dubai day. The bars cover the Claude models only for now; voice and the judge stay in the tiles. Nova's own builder wrote it, and Claude reviewed it.

A DAY button in the Chat page's title row (on the laptop window) opens a day chooser. The day is read back from Nova's own records, a page at a time, and presentation mode hides every word. Live since 24 October.

Since 20 October 08:27 the backend reads the backup tool's own record of the last run (times and counts only) and raises one alert when no good backup came for 50 hours or the last one needs a look. Nova's own builder wrote it, and Claude reviewed it. The safety check's verdict and muted alerts are not on the page yet.

Since 20 October 11:41 the backend reads the same list as the stats panel inside Claude Code, at most 60 entries with the open ones kept first. In presentation mode every text turns into dots. Nova's own builder wrote it, and Claude reviewed it.

The page is built from his calendar and Nova's own list of tasks and reminders. Mail shows only as counts, never a word of a mail. H brings Home back from anywhere, and presentation mode turns the private parts into dots. Live since 26 October.

The brief is filled in from a fixed template on the laptop, so it needs no AI at all, only the spoken voice. It stays quiet in presentation mode, in a meeting and while recording. Live since 27 October.

The overview keeps the Nova look and shows only what matters first. Presentation mode masks it. Live since 26 October.

Each state keeps a slow motion of its own in every power mode, and only a hidden window rests. A drawing error in one small part can no longer stop the whole animation. Live since 25 and 27 October.

Nova Home runs as the tablet's only app, with a pull-down quick panel. Every page works upright and sideways, with bigger touch targets in touch mode, and the keyboard lifts only the chat line, never the whole page. Live since 22 October.

A quick-panel switch, "Keep Nova connected while locked", is on by default (it uses a little more battery). A locked tablet gets only what changed, which saves data. Live since 26 and 27 October.

The switch is on the laptop only (Settings, Devices). Live since 25 October.

Fixed rules decide what is slipping, with no AI and no cost. Most rows have a one-tap fix, presentation mode hides their words, and Telegram gets at most four of these lines a day, only in the daytime. Live since 28 October.

One store on the laptop holds every task. A wide screen shows four columns and an upright tablet a list grouped by step; a tap on a task's step moves it. Nova's own suggestions wait for his Start. Live since 28 October.

The backend reads the projects' history slowly and read-only, keeps 90 days of it, and shows titles, numbers and times, never code. Live since 28 October.

Nova suggests clients and Aman keeps or skips each one. A draft is one short AI answer or a fixed template, and it leaves the laptop only when Aman copies it or opens it in his own mail; nothing is sent by itself. Live since 28 October.

Nova Home checks every few seconds that the page still answers, without reading anything on it, and never touches a page that answers. The laptop's Reload came on 28 October, and Nova Home 1.0.8 went onto Aman's tablet the same evening.

The laptop records only while Nova's listener is active. When he stops talking (about a second of quiet), only the part with speech goes to OpenAI's speech-to-text. It understands English, Hindi and Gujarati, and knows words like Nova, n8n and Telegram. Pressing again sends at once, Esc stops, and pressing while Nova talks interrupts it. The limit is 150 voice messages a day.

A small keyword engine runs on the laptop and listens only for its fixed phrases. No sound and no words are sent anywhere or kept, only daily counts. It does not wake while Nova is in a Google Meet, and it pauses when the screen is locked or another program holds the microphone.

The MIC badge follows the real state of Nova's microphone, and a click or key M switches it. When "Hey Nova" is heard, the bar shows it. Help's "What you can say" lists phrases such as "Nova, next slide", "mute", "stop listening" and "cancel".

The same on-laptop keyword engine knows twelve fixed phrases and presses the slide key or changes the volume directly. They cost nothing and never reach the brain.

Nova speaks the greeting before it starts recording, so it never hears itself. Each spoken request asks for a warm, natural style: one short line while it works, then the answer in one to three sentences. The full answer stays on screen.

Shortly after Nova's last word the microphone listens again. A follow-up needs at least one second of real speech. Quiet costs nothing, because nothing is sent. Esc or the button ends it.

The brain knows the microphone hears the whole room. For words not said to Nova it answers with a fixed marker, so the voice says nothing and the conversation ends. On the real brain, pieces of a phone call were ignored and real requests were answered.

A filter turns the written answer into speakable text: links become "a link", file names "a file", and emoji and layout marks are removed. The answer is cut at a sentence with "The rest is on the screen."

Nova's answer streams to ElevenLabs' fast model and plays on the laptop as it arrives. The chosen voice is pinned in Nova's settings, so it stays the same after restarts, and the same voice is used in Google Meet.

The Settings page reads the available voices (no key is ever shown) and saves the choice on the laptop. Nova uses it from the next spoken answer.

Each line is made once per voice and kept as a sound file on the laptop. A background helper makes any missing lines while Nova is idle. The Usage page counts the lines "played from this laptop, no credits".

When an approval card appears for a spoken request, the kept line plays once per card. Nova never reads the command itself aloud. Questions get a lead-in: "I need your authorization, sir: ..." or "I need your input, sir: ...". Polite offers like "Let me know if..." are not spoken.

The wake word runs on the tablet itself. A conversation ends on End, a pause of silence or a locked screen, and it is at most ten minutes or twenty requests. Live since 23 and 24 October.

The check was tuned to his real voice on 24 October and learns from his Telegram yeses since 25 October.

It is designed on OpenAI's real-time voice model, with a LIVE badge before any sound leaves the laptop and a daily time limit. Four of its ten pieces are built. Aman paused it until he chooses to finish it.

Live voice joined with Nova's browser and desktop control, which already work with typed and voice messages.

It comes with live voice. Its place on the page bar (Ctrl+7) is kept free for it.

The key is set in a short sitting with Aman. The tuning uses a two-minute recording of his slide phrases. Today, key M works in the Nova window.

Aman chose on 19 October: when he talks to Nova. Since then Nova's closing words work in Hindi and Gujarati, the window counts which languages his voice messages are in, and Nova's voice is told when an answer is in Hindi or Gujarati. Next comes his 10-minute test, then each language is measured. Meeting notes are English-first today (about 21% character errors on Hindi tests); a Hindi model that was tested did far better.

Aman said yes on 28 October, and the laptop and tablet sides are being built.

Nova opens a separate meeting window that only it controls, signed in to Nova's own Google account, with its own virtual microphone and speaker. The laptop's real microphones never reach the meeting, and the meeting's sound never plays on the laptop's speakers. It leaves when asked, when the meeting ends or it is removed, after 5 minutes alone, after 15 minutes without captions, or after 3 hours.

The meeting window draws Nova's camera itself in 720p HD at about 25 frames a second, and the backend changes its state as Nova listens, thinks and speaks. Nobody in the meeting needs to install anything.

The arrival animation still plays on its camera; the notes, the ear and the captions start as before. A call by name is answered only when the judge is sure it was said to Nova. Tell the guests yourself that Nova takes notes.

Nova reads Google Meet's live captions four times a second and collects the speaker's words after its name until the sentence is complete. It also catches common caption mistakes, like "Novelo" or "No". The answer comes from Nova's brain, in a conversation of its own for that meeting, and is spoken in its ElevenLabs voice. Nova unmutes itself before speaking. First words are ready in about 2 seconds.

Simple rules first pick lines that might be for Nova. A small, fast judge model then reads the recent conversation and decides in about 0.3 seconds whether the line is for Nova, about Nova, for others, or unfinished. Only sure cases get an answer. It costs about 6 cents per 1,000 decisions.

A keyword engine on the laptop listens to the meeting's sound for "Nova" and matches each find to the person whose caption was moving at that moment. Nothing of the sound goes to any outside service.

Nova opens the file in a hidden tab of its own meeting window and shares only that tab, never the whole screen or another window. A PDF becomes one picture per page in Nova's own dark slide viewer. Slide words act without the AI, but only while Nova presents or when its name is said.

It is a web presentation kept on the laptop, with smooth animation and auto-play. The narration was recorded once in Nova's ElevenLabs voice, so it costs no extra AI call. While it plays, Nova ignores its own name in the narration.

During the meeting Nova records its sound to a private file on the laptop. Afterwards a local speech-to-text model and a speaker-separation model write the transcript and name each voice from the captions. If that fails, the captions are used. An hour of meeting takes about 20 to 40 minutes in the background. A real one-hour meeting gave 111 lines of notes.

Nova's brain writes the summary from the transcript and sends it to Aman's Telegram. If the laptop restarts first, the next start finishes and sends it. A summary is made only when others spoke or Nova answered someone.

For each request under his name, Nova compares the voice of that line with samples of his voice, kept only as numbers, never sound. It runs while the judge decides, so it adds no wait. A request whose voice does not match is not carried out. Nova holds it and asks Aman on Telegram instead (see the next feature). One honest limit today: Aman's voice through a meeting still scores like a stranger's, so his own orders in a meeting are held too. Tuning that is planned (see "Knowing Aman's voice better in meetings" below).

Each check has a short code, closes after 10 minutes or when the meeting ends, and is limited to 5 per meeting. Nova never says in the meeting why it is checking or who may give orders. Aman answers with one tap (see "One-tap Yes on meeting checks" below).

Guests' words (and words under Aman's name without his voice) run in a separate, guarded conversation. A safety add-on in Nova's brain blocks every tool there except web search, and this was proven on the real brain. Guests' words are treated as information, never as instructions.

After leaving, Nova waits 10 minutes before writing the final notes. A rejoin of the same meeting stops that wait and carries on with the same notes and conversation, adding a "Nova rejoined at ..." line.

Nova's attention is a small set of states (idle, addressed, thinking, speaking, invited, engaged, yielded). Its ear hears when people start and stop talking, and lines that are not clearly for Nova need a surer score, above all right after Nova has spoken. Slide commands stay instant, a slow answer gets "One moment.", and lines Nova kept quiet on still go into the summary. Aman asked for this on 18 October ("more proper and advanced"), and the settings come from the meetings of 16 to 18 October. Live since 19 October.

Nova counts who is in the meeting and checks that the voice is Aman's. As soon as anyone else is there, it speaks only when called by name again. Live since 25 October.

The tap goes through Nova's Telegram button kit straight to the meeting service. The check's code comes from the card itself, not from the AI, and a card older than 10 minutes answers "closed".

The backend recognises the meeting's own conversations and the notes job by their names, and gives them these titles and a "meeting" channel.

A short test meeting with him, then one of three fixes: fresh samples, samples taken from a meeting, or new thresholds. Nothing changes before he chooses.

Nova turns a .pptx, .docx or .xlsx file (at most 50 MB) into a PDF first inside the document box, the same sealed box the document tool uses, so nothing is installed on the laptop: LibreOffice lives only in that box. It takes up to two minutes the first time; the pages are kept, so the same file starts at once the next time. An older format (.ppt, .doc, .odp) must be saved as .pptx, .docx or a PDF first. Live since 20 October 05:49, tested with made-up files; it has not been used in a real meeting yet.

It works on the laptop window, for Google Meet links on Aman's own calendar. A restart of Nova forgets the plan, and the card says so. Live since 27 October.

The calendar page and feed are live, but joining by itself is not designed yet.

Research found that Meet's audio carries each speaker's identity, which some meeting bots read. Not designed for Nova yet.

On the meeting plan's "later" list, not designed in detail.

The same short summary run that writes the meeting notes also lists the action items, at almost no extra cost. Nothing becomes a task without his Keep, and Telegram gets one line per meeting. Live since 28 October.

This is the brain's own Telegram channel, locked to Aman's account. Slide commands such as /next run instantly, with no AI. Two small local additions make Nova's own tap buttons work.

Nova built this kit itself, and it runs inside the Telegram connection. Every tap is checked (only Aman counts), recorded, and passed to the right job.

Nova uses the brain's own WhatsApp channel on a spare number Aman gave it. Strangers get no answer, and groups are ignored.

Every channel goes through the one brain. A send command pushes a result to Telegram or WhatsApp, and the dashboard also shows work started on other channels.

The Chat page lists the 20 newest messages, each with its first words and its channel. Home shows the newest in its bottom strip. Presentation mode hides them.

The full messages come from Nova's own records, 50 at a time with "Show earlier". The view is read only: Aman writes on the phone, or chooses HERE to type, and pressing / or starting a voice message brings HERE back by itself. Nova's answers keep their cards and tables, a picture shows as "[attachment]", and presentation mode hides every word.

Re-apply tools put Nova's small changes back after each update, in a fixed order, and can check and undo them. Each is tested on a copy first.

A small helper on the laptop takes one action at a time from Nova (open an app, open a web address, run a web search) and puts the window where he asked. It works from Telegram, the dashboard and by voice.

A Nova add-on inside the desktop acts on one window at a time: bring it to the front, move it to another screen, place or resize it, full screen, minimise, or close it politely (the app can still ask to save first). It never reads what is inside the window.

The brain's own desktop control looks at the screen, then moves a pointer, clicks, types and presses keys. A Nova cursor arrow shows on screen while it works. It works from the dashboard, Telegram and WhatsApp.

Nova presses the right key (arrow, Home, End, Escape) in the window in front, through the desktop helper. It works with "Hey Nova" too.

The desktop helper takes one picture of one screen or one window. Nova can also look at the picture to answer "what's on my screen?" It never takes a picture by itself, only when Aman asks.

Nova drives a real Chrome window on screen, signed in only to Nova's own Google account, never to Aman's. It works from the dashboard, Telegram and WhatsApp, so from the phone it can look things up in a real browser too.

n8n checks the inboxes every 2 minutes. Each new mail is sorted by an AI writer on Claude Haiku that has no tools at all, so a trick written inside a mail cannot make Nova do anything. Then the card goes to his Telegram.

The page reads the mail watch's own records on the laptop. It never draws a mail's own web code and never opens a link by itself, and nothing on the page sends a mail. Presentation mode hides it. Live since 25 to 27 October.

The mail watch adds it as it sorts the mail, at most ten a day, and the Undo takes it off again. Live since 26 October.

An n8n automation reads both Google calendars, puts the events in time order and sends them to his Telegram.

The backend reads his calendars on the laptop, read only, through an n8n feed. Clicking an event shows the video call, place, organizer, guests and notes, and "Join in browser" opens the meeting. Presentation mode hides titles and names.

The tap goes to an n8n automation on the laptop, which asks Nova's writer for the reply text and saves the draft or the calendar entry. Nothing is sent, paid or deleted. It is switched on but not yet proven end to end.

The daily check reads a short list of sign-ins and their end days: the Google sign-ins are warned on 23 and 24 October, before they run out on the 25th. Every 10 minutes the backend also asks n8n for failed runs that say "needs to be reconnected" or "expired", without changing anything in n8n. Aman is also publishing Nova's Google app, so Google sign-ins stop ending every 7 days.

An n8n automation picks the time and gathers the stories. Nova's writer, which runs on Claude Sonnet and has no tools, writes the post from the story and the article. A Telegram button add-on that Nova built itself sends his taps back to n8n. Nothing is posted without his Approve tap.

Nova runs an n8n automation that holds the X connection, so Aman's X account details never pass through a chat.

An n8n automation holds the LinkedIn connection and is run once, by hand, for each approved text.

An n8n automation reads a tech news feed and writes up to four drafts a day. The dashboard reads them, read only, and shows the time, slot, text and source.

A scheduled job only reads Hacker News and writes the queue; it never posts. The dashboard lists the stories highest score first, with rank, points and status. Paused since 27 October (Aman's choice); it comes back on his word.

Nova writes one line per post and the Social page reads it: newest first, sent or failed (with X's reason), and counts per platform. Presentation mode hides the texts.

A scheduled job sends the queue as cards through Nova's Telegram button kit. It only sends cards, never posts. It was fixed on 18 October and runs every 6 hours (12 runs of the job since 17 October by 19 October 21:34, the last one ok). Paused since 27 October (Aman's choice); it comes back on his word.

A tool reads n8n's past runs, counts only real taps (not tests), and writes each post into the log the Social page reads. It runs by itself every night at 03:00, before the backup, so a tap post is in the list the next morning. To show one at once, run it by hand right after the tap.

The cards read the same records as before; approving still happens on Telegram. Live since 27 October.

The X automation reads the trends once a day and keeps them; the page asks for them at most every ten minutes. Live since 27 October.

A draft waits for his yes on his phone or on the dashboard, and Approve asks once more before anything is posted. Live since 28 October.

Built like the X button route in n8n. Aman chooses between the button route and the full outbox.

Only a switched-off sketch exists. Each read costs X credit, so it waits for Aman's yes and a daily cap.

Nova uses web search and its browser, links every source, and saves the report in its work folder. Bigger research questions go to the stronger Sonnet model.

Nova looks them up on the web, reads Aman's list of offers on the laptop, and writes a one-page brief. A meeting card from his mail also has a "Brief me" button.

Nova uses web search and its own browser.

Nova saves each note on the laptop with a short number, so Aman can close, reopen or delete it later. The weekly review reads them back as a summary.

The brain has its own scheduler. Nova creates a timed job, and the message goes out when the time comes. It works from every channel. Since 22 October reminders kept in Nova's own calendar also reach Telegram on time.

A timer on the laptop runs the check every hour during the day until that day's check is done. It only reads Nova's own lists. It also says when Nova's post history was not brought up to date for two nights.

Nova keeps the list on the laptop and saves a change only after Aman says yes. The daily check reads it and warns before each renewal.

Each skill is a written recipe Nova reads when a request fits. Some need an account Aman has not connected yet (Notion, Airtable, GitHub), and the Apple ones don't work on this Ubuntu laptop. Most were not tried one by one here.

Everything sits in one small database on the laptop, the Nova store, which the dashboard and the tablet share. Live since 22 October.

Words written in a picture cannot make Nova do anything else. A picture that looks private is not read, and its photo is removed at once. Live since 26 October.

A full-screen pad with palm rejection, an eraser and undo. Nova reads the handwriting as it reads a photo. Live since 26 October.

Nova never sends one. A mail leaves only through his Copy or Open in mail, and a WhatsApp message only after his Yes on a card, once. Nova writes the drafts without seeing anyone's number or address, and none of it is used in a meeting or a scheduled job. Live since 28 October.

The search and its index stay on the laptop and cost nothing to run. Aman set it up on 28 October and is trying it out; it stays marked as being built until he calls it done.

The skills are installed. The timed n8n watch is switched on once Aman gives his two lists (sites, and subscriptions with dates and prices).

An n8n automation uses a travel price service (free up to 250 searches a month), with one tap before each search. It waits for Aman to open the free account and enter its key himself.

It needs the document tool (see Coming next) and his template.

Each would reuse what Nova already has (search, notes, desktop control) plus n8n's clock. None is designed yet.

Every dashboard request reminds Nova to keep its list, and the dashboard reads the list every two seconds. The item in progress is lit, done items get a check, and cancelled ones are struck through. The list survives even when the brain shortens a long conversation.

Every five seconds the backend reads the brain's list of background jobs: what runs now, what ended in the last 24 hours and, for a failure, its last line. Nova is taught to look at a job before saying it is running, and to answer "how is it going?" from the job itself.

One helper runs the build in a copy of the project and logs its progress (see the next feature for whose key pays). The build shows under background work, and Aman gets a Telegram message when it ends. A finished copy goes live only after it is checked and Aman says yes. Since 27 October the development of Nova itself is Claude's work; Nova builds only what Aman asks it for himself, and manages the rest.

Nova has n8n's own tools and a skill that teaches it to build and test safely: sending steps are faked during tests, and a workflow that posts stays off until Aman's yes.

Every 10 minutes the backend reads n8n's list of failed runs (the automation's name and id, never what the run held). Each automation is told at most once a Dubai day, and the message begins "From the laptop (not a reply):" like everything the laptop sends. Live since 20 October 02:15. Nova's own builder wrote it, and Claude reviewed it before it went in.

A helper starts a separate run of Nova's own brain on Claude Opus in a copy of the project, with a time limit for each build. It shows in the Task box as a Nova build.

Nothing is ever forced, and nothing is merged without him. Live since 24 and 25 October.

The backend looks at Vercel and GitHub every few minutes, at most twelve lines a day. Live since 25 October.

The site is a Next.js project on Vercel, at nova.primexmeta.com since 29 October. Every change goes through a pull request and goes live when Aman merges it. A Sunday update is built and waits for Aman's word to switch it on: Nova would write the week's changes as an article, two privacy checks read it, and it goes live only when Aman taps Publish. Live since 25 October.

The backend reads token counts from the brain's records and prices them from one price list covering Haiku, Sonnet and Opus. A bar shows where today's cost goes, and each model has a card with calls, tokens, cache use and roughly what one call cost. Every figure is marked EST (estimate), and a part that is not counted says so instead of showing zero. Aman has also set spend limits in the Claude and OpenAI consoles.

Nova counts each spoken piece from the provider's own numbers and books it with a price list kept on the laptop. Roughly: OpenAI speech about 2 cents a minute, ElevenLabs about half a credit per character.

Every judge call is counted from the provider's own token numbers (never the words) and priced. It appears as "Judge (TypeSafe)" and is included in the totals.

The count comes from the same records as the Usage page, priced the same way, and builds keep limits of their own. Aman gets one Telegram line the first time it is reached on a day. Live since 25 October.

It joins the estimate of Claude Code's work on the laptop with what Nova's own builds used. Live since 24 and 25 October.

Telegram is locked to his account and WhatsApp ignores unknown senders. Nova's persona says only Aman's own messages are requests, and text from web pages, mails or files is treated as a stranger's words.

The rule is written into Nova's own instructions on every channel, and a yes in the dashboard counts as much as one on Telegram. Since 25 October a "Yes, once" / "No" card also holds those steps when Nova uses its browser or the desktop.

Every service listens only on the laptop itself, and the firewall blocks incoming traffic. Presentation mode hides private text and money while recording.

One tool takes it without showing it on screen. Snapshots and backups are checked so that nothing private is copied into them.

The stop command ends every running request, pauses the brain and stops the services. The microphone command remembers "off", and nothing spoken can switch it back on.

A timer runs the backup at about 03:30. The copy of 18 October held 13,869 files and 27 databases (about 308 MB) and was complete and verified in 13 seconds. Since that day it also holds Nova's work folder (its Telegram kit and its writer for n8n) and the dashboard's own state.

The snapshots are a local history, made through one tool that first checks that nothing private is inside. They are never pushed anywhere.

The check tools were rewritten in a copy and tested against the live system before they came in, on 18 October.

It looks every few minutes, tells a thing only after seeing it twice, and writes under its own heading, never inside Nova's conversation. Nova's own builder wrote it; Claude checked it and switched it on.

The backend reads the brain's own record afterwards, so Nova is not slowed down. It keeps the time, the chat, plain words, the site or app and the button's name for 30 days; never what was typed, a page's text or an address's path. In presentation mode the band leaves the site or app out.

Live since 25 October.

A small recorder of its own uses the desktop's built-in screen capture, so no pop-up appears and the desktop's red recording dot shows. The graphics chip encodes the video, using about 2 to 3% of one processor core, and the video is saved in one-minute parts so a crash can't break the file. Files stay on the laptop's big disk, named by date. Nothing is uploaded.

The recorder takes the microphone and the laptop's sound on the same clock as the picture. "Hey Nova" keeps working during a recording, and the meeting's sound devices are never used.

Recording switches presentation mode on and holds it until the recording stops; Aman can turn that off on purpose for one recording. It only masks the Nova window, and Nova warns him when he records a whole screen where other windows show.

The marker is placed at the moment "Hey Nova" was heard, not when the command finishes. Chapter times follow the cuts and speed-ups, and YouTube's chapter rules are applied (first chapter at 0:00, at least three chapters, at least 10 seconds apart).

The camera is recorded to its own file and placed in the corner at export, so it can still be moved or dropped later. The first real recording with the camera (19 October) came out black while the screen and the sound were fine; the cause is being checked (a closed lens cover is the first guess).

Captions come from a speech model running on the laptop, so no sound is sent out, and Aman reads them before uploading. The big screen exports at 1440p, which YouTube streams at higher quality. A 10-minute video takes about 5 to 6 minutes to export.

Recording only the Nova window works since 24 October. The key is set in the same sitting as the microphone key. The silent phrases come only after a test of false wake-ups with Aman's own voice.

n8n's own YouTube connector, with the channel signed in by Aman himself. Until then he uploads in YouTube Studio with the files Nova prepared.

Aman runs a small install command himself. Editor caches go to the big disk.