Chat (NeuralSpark agent) #
Everyday questions, writing, files and research in one conversation, with no project set up first.
Open Chat in the sidebar. A new conversation shows "Explore an idea", "Create something" and "Get work done", then "Pick up where you left off". The header menu offers Rename, "Pin conversation", "Save a copy of this conversation", "Save as Markdown", "Archive conversation" and "Delete conversation…". Under each answer: Copy, "Good response", "Bad response", "Branch into a new conversation" and the time.
What stands out: Extra tool families (cards, design, PDF, Office, Google, mail, browser, automations, connectors, NeuralVault) load only when the work needs them.
Limits: No Git tools and no stateful project shell by design. Coding in a project is pointed to Code.
Deep research #
A longer, sourced report when a quick answer is not enough.
Pick Deep research in the Agents list before starting. It is described as "Thorough research with sources: many searches, cross-checked figures and a structured report." Its instructions allow up to twenty searches and thirty page reads, prefer primary sources, and ask for numbered sources.
What stands out: It keeps the cards family loaded from the first step, as Designer keeps the design family.
Limits: It costs more of the person's own provider searches and takes longer. The agent description says so in plain words.
Designer #
A conversation kept on one deck, report or page until it is finished.
Pick Designer in the Agents list. It starts each piece with design_start, follows the design guide that returns, and exports as PDF, PowerPoint or Word through the design tools.
What stands out: The whole conversation stays about one designed piece, rather than design being one answer inside a general chat.
Limits: Design work is also available in Chat. No proof was found in the opened code that it can work without a connected model.
The message box (composer) #
The one place to type, attach files, choose a model, set file access and send work, in Chat, Code and Cowork alike.
The placeholder reads "Ask anything…". The plus button opens "Add files or run commands" with File, Goal, Plan, Feedback, "Summarise earlier messages", File access, Model and Export. A message typed while NeuralSpark works waits as a queued message with "Send now". The setting is "While NeuralSpark is working" with "Queue my message" or "Send now". Ctrl+Enter does the opposite.
What stands out: The same composer carries file access and model choice everywhere, chosen from the agent, not from a separate app.
Limits: With no model chosen, it is blocked with "Choose a model in Settings › Models to start". A conversation open in another window cannot be sent to from here.
File access modes: Read only, Can edit files, Full access #
Per conversation, whether NeuralSpark may only look, may change files in the workspace, or may take fewer confirmation steps. See also Safety and control.
Pick the mode beside the message box, or set the starting mode at Settings, General, under File access: "What NeuralSpark may change in your workspace in new conversations". In plan mode the sandbox resolves as read-only whatever the usual mode.
What stands out: The mode is enforced outside the model's instructions. A write outside the workspace in Can edit files mode is refused with "Not run: it would change a file outside this conversation's folder."
Limits: Full access never asks, so tools set to Ask in a connector still run. The person's own terminal command tool is refused in Full access because it must always ask.
Plan first #
NeuralSpark sets out the steps and waits for approval before anything changes.
Press "Plan first" beside the message box. Its title reads "NeuralSpark plans the work first and asks before it changes anything". While planning, other tools are held with "Plan mode is on, so … was not run: nothing changes until the person approves the plan." The plan shows above the message box and ticks steps off. Ctrl+Shift+E switches Normal and Detailed views.
What stands out: Files and commands are held by the sandbox as read-only, and the guard also holds tools the sandbox cannot see, such as remote computers, Git, designs, Google, PDFs, automations, the browser and connectors.
Limits: Reading, searching and looking stay allowed while planning, so a plan can still cost model steps.
File, picture and audio input #
Hand over the document, picture or recording the work is about, instead of describing it.
Press plus, then File, or drop the file on the message box. Accepted recordings: MP3, M4A, WAV, WebM and OGG. A recording is transcribed first, with "Transcribing" plus the file name and length above the box, then the transcript is attached as a text file. Recordings up to three hours are accepted for MP3 and WAV. A very large pasted block is saved word for word under the conversation's `.neuralspark/work/pasted` folder and the model is told the path.
What stands out: Pictures go to the model as pictures where the model can take them. The Muse Code route reads up to eight pictures from a message. Pasted figures are calculated by a script over the saved file, not summed from part of the text.
Limits: This PC's own speech model hears WAV only. Other kinds need the person's OpenAI or Gemini key, billed by the minute. The recording itself is never kept.
Private chat #
Ask something without it being kept. See also Safety and control.
Use plus, then Private chat, or Ctrl+Shift+N. Above the box: "Private not saved, no memory". End finishes it.
What stands out: It is a mode of the normal chat, not a separate app, so the same model and tools apply while nothing is kept.
Limits: Not saved, not in Search, the sidebar, Library or "Pick up where you left off". Removed when another conversation opens, End is pressed, or the app closes.
Today card and Morning brief #
The day starts with the few things that need you first.
The Today card at the top of Chat home shows up to five things with one button each, such as Review, Fix checks, Meeting note, Open and Run again. "Not useful" shows fewer of that kind later. In Settings, General, the Morning brief switch sets its time and "Write a short summary with AI".
What stands out: One button per item acts on it from the card, and the brief learns from "Not useful".
Limits: The card hides when there is nothing for today. The AI summary appears only when "Write a short summary with AI" is ticked. If the app was closed at the brief time, the brief is made when it next opens.
Memory #
Lasting facts carry between chats without retyping.
In Settings, Personalisation, then Memory: topics with General or workspace, "From a conversation", Show, "Use general memory", edit, forget, "Remember things between conversations", "Keep memory in NeuralVault" and "Forget everything".
What stands out: Memory can be kept in NeuralVault, and single topics can be shown, edited or forgotten rather than only wiped whole.
Limits: Passwords, keys and card numbers are never kept.
Custom instructions #
New chats start the way you like.
In Settings, Personalisation, set Custom instructions, "Ask before closing", and the "New tasks show" switches for starting points, recent chats, playbooks and team work.
What stands out: The starting points on a new task can be switched per person, so the home shows only the ways in that person uses.
Limits: "Ask before closing" shows "Close NeuralSpark" with "Dont ask again" to stop it. No wider limit is stated in the guide lines opened.
Output styles #
Answers come in the length and depth you picked.
Under Personalisation, "How answers are written" lists Balanced, Concise, Explanatory and Tutor in the progress records, kept per conversation.
What stands out: The style is kept per conversation, so a teaching chat and a short answer chat can run side by side.
Limits: The guide section itself states that exact screen steps need the Personalisation page. That page body was not in the section proofs, so no screen steps beyond this are claimed here.
Bring your history #
Old chats from other apps become searchable here.
In Settings, Personalisation, then Bring your history, use "Bring your history" for a ChatGPT, Claude or Gemini export ZIP, then open results in the Imported workspace.
What stands out: Imported chats land in their own workspace, apart from current work, and can be searched there.
Limits: The same export again adds only what is new.
Sidebar, conversation dots and the Notices bell #
You can see what needs you without opening each chat.
The switch at the top reads Spark, Vault and Bot, then Chat, Code, Cowork, Library, Agents and skills, and Automations, with Settings at the bottom. A dot beside each title pulses orange while working, shows an orange ring while waiting, green when finished, red when stopped. Hovering says Working, Waiting for you, Finished or Stopped. Drag the top rows, workspaces and conversations into your order. The bell beside the account lists finished tasks, approvals, questions and automations for a week, with "Mark all read" and Clear. A notice opens its conversation.
What stands out: Automations shows NeuralBot items among them, and one bell holds Spark and Vault news together (see also This PC, Notifications).
Limits: Opening a conversation clears green and red dots only. The bell list covers a week.
Finding a conversation and Search all conversations #
Find words across past work and open the result at that message.
Press Ctrl+Shift+K or the sidebar search, then "Search every word". Filter by All, Chat, Code or Cowork, workspace and time, then open a result at that message. Tab takes up a grey suggested next message in the empty box. Ctrl+semicolon opens "Ask a side question". Ctrl+K opens the Command palette. Ctrl+slash lists shortcuts.
What stands out: A result opens at the message, not at the top of a long conversation.
Limits: Archived results unarchive as they open when "Include archived" is on. A suggestion sends nothing until you send.
Diagrams, maths and file cards in answers #
A process reads as a picture, maths reads as maths, and made files open where you are.
A mermaid block is drawn with "Show code", "Show diagram", Copy and "Save as picture". TeX in dollar signs is typeset. Click a file card to preview in the side panel. Open opens it in its program. Several files made together sit under one heading with a count. An email file shows "Draft not sent" with Open to send.
What stands out: The same answer can hold a drawn diagram, typeset maths and the finished files, each usable where it sits.
Limits: A diagram that cannot be drawn stays as code with "This diagram could not be drawn." Nothing is sent until you press Send in Outlook for email files.
Short videos #
A short clip for a post from a sentence.
With Gemini or xAI connected, ask for a video, approve the length and price card, wait on "Making the video", then use "Save video" or "Use in a post".
What stands out: The price is shown and approved before the clip is made, on your own key.
Limits: 4 to 8 seconds, billed on your own key. The guide says OpenAI, Anthropic, Muse, DeepSeek and others make no videos through NeuralSpark.
Workbook in the side panel #
A spreadsheet edits where you read the answer.
Click a cell and type. Enter saves down, double click or F2 edits, Delete clears, Ctrl+Z undoes. "Open in Excel" opens the real file.
What stands out: A small fix is made on the shown sheet, not by making a new file and hunting for it.
Limits: A table filter filters the view only.
Word document or deck in the side panel #
You fix words on the page, not in a new file.
Select words for Rewrite, Shorten, "Make it friendlier" and Ask. Review in the status bar offers Keep, Undo, "Keep all" and "Undo all". Ask NeuralSpark to answer on comments. Present or Presenter view shows decks.
What stands out: Each change can be kept or undone on the page, one by one or all together.
Limits: Word changes are tracked changes. A PDF offers Ask only.
Code home and Tasks #
Coding jobs run apart so several can go at once, each in its own copy of the project.
The Code home shows Tasks marked Running, "Needs you", "Ready to review", "No changes yet" or Merged, with "Merge into the branch", "Pull request" and a menu for "Update from the branch", "Copy folder path" and "Remove task copy". Below: Your projects with "Own copy for each task", "Add a project", and "Use NeuralSpark from your editor (Zed, JetBrains, VS Code)".
What stands out: A task in a Git project works in its own copy on its own branch. Code also keeps a code lookup tool for definitions, references, implementations and types in TypeScript, JavaScript and Python.
Limits: No Google, design or card tools by design. A JavaScript task copy starts without the project's packages, so tests cannot run until the copy is prepared.
The code editor #
Read and fix a file yourself beside the task, without opening a separate program.
A file opens in the editor beside the task. Save or Ctrl+S saves. Ctrl+Shift+F searches every file. Ctrl+P opens "Go to a file by name". F12 goes to a definition, Ctrl+F12 to an implementation, Ctrl+click does the same. If the file changed on disk, saving asks "Keep mine and save" or "Load theirs".
What stands out: It is the editor VS Code uses, loaded in its own frame so its styles never meet the rest of the app. It follows the task's edits on disk and warns where unsaved changes would clash.
Limits: Only a folder open in NeuralSpark is served, and no path may leave that folder. Colouring comes from the editor's tokenizers. Definitions and types come from code lookup.
Code tools: Go to file, Search, Changes, Tests, Problems, Pull requests #
File search, the change list, test runs and type problems in one panel beside the task.
The Code tools button (Ctrl+Shift+F) gives "Go to file", Search, Changes, Tests, Problems, "Pull requests" and Project. Search runs with ripgrep where present and falls back to reading files. Tests offers "Run all", "Run failed" and Coverage. A failure shows "Failed at" with a file and line. "Open the test" opens it there. "Ask the task to fix" names the file and line in the message box. The runner detects Vitest, Jest, Mocha, Node's own runner, pytest, Go, Cargo and .NET. Problems comes from TypeScript's checker or Python's pyright, both shipped.
What stands out: Test results are read from each framework's own report, not guessed from printed output, and each failure is tied to the project's file and line. Where a program is missing, the panel says so instead of reporting a pass.
Limits: A run that reports no tests and does not finish cleanly is shown as "Did not run", never as all passed. WSL project folders are not run inside Linux yet.
Review, diffs and Check my changes #
Every change in the task, line by line, with an option to have NeuralSpark read the changes for problems before you keep them.
Review lists every change, each line numbered, with "Wrap lines". "Check my changes" leaves comments on lines, each with Fix or Dismiss, then "Check again". "Put back this file" and "Put back this part" ask first, and Undo takes it back. The reviewer is told to comment only on bugs, wrong behaviour, security and privacy problems, data loss, unhandled errors, races and missing tests for risky logic, with no style points and no praise.
What stands out: Three levels: Quick, Standard and Deep. Deep also reads the changed files whole and checks its own comments in a second pass. The diff is taken from where the task branch and its source branch last met.
Limits: A diff longer than the level's limit is cut, and the review says it reviewed what was shown. At most 12 comments at Standard level, most important first.
Terminal #
Run commands in the task's folder, and let NeuralSpark read what your terminal shows without copying output into the chat.
In Code, the Terminal button or Ctrl+grave opens a terminal in the side panel in the task's folder. In any conversation it opens from the side panel's New tab, then Tools, then Terminal. The arrow beside "New terminal" chooses Command Prompt, PowerShell or another installed shell. NeuralSpark can read what the conversation's terminals show when asked, and run one command only after you allow it, showing the command, with "Allow once", never remembered.
What stands out: The read tool takes the terminal screen directly as plain text with keys and tokens masked. Words like "the right terminal" are treated as where the panel is, never as a terminal name.
Limits: A command is typed only when the terminal sits at its prompt, never into a running command or over words you started typing. It never reads another conversation's terminals.
Cowork #
A longer piece of business work handed to a small team that plans it, shares the parts out, checks the results and delivers finished files in your folder.
The Cowork home shows Your team's work with tasks marked Working, "Waiting for you", "Could not finish" or Finished, and Playbooks below. In a task, Files shows everything made. Team shows Lead and each helper as Working, Finished, Starting or Stopped, plus The plan as "Not started", "In progress" or Done. The lead works in four steps: Plan, Delegate, Check and Deliver. Specialists are named researcher, writer, designer and analyst.
What stands out: Choosing Cowork is itself taken as the request for a team, so teammates are created without asking again. One document has one maker.
Limits: Nothing is sent, submitted, published, paid for or deleted without asking first. Draft emails and messages are made as files or drafts, never sent by the team itself. Team tools exist only while a conversation is in Cowork.
Playbooks #
Common business jobs with a ready-made shape: details to fill in, steps, team and checks on each file, so you fill a short form instead of briefing from scratch.
Four playbooks show on the Cowork home. "Search playbooks" searches all of them and "Browse all" opens Playbooks with categories including Compliance and admin, Finance and bookkeeping, Marketing, Operations, People and HR, Reporting and strategy, Research, and Sales and customers. A playbook opens its short form. Start writes the request into the message box. There are 50 playbook files, from award-rate-check.json to win-back-campaign.json. Month-end bookkeeping prep asks for Business name, Month, Bank statement export and sales and expense records, works out GST collected and paid, and delivers a workbook and bookkeeper notes.
What stands out: Each deliverable carries its own checks, such as "the price matches what was given" or every bank line appears once, and matched plus unmatched equals the statement total. A playbook file that fails its own shape checks is left out, never half used.
Limits: Only researcher, writer, designer and analyst may be in a playbook team. A playbook is a brief, not a finished file: a missing bank statement or records file stops that job. If no playbook fits, the agent plans the job itself.
Run this as a team #
A Chat request that has grown into several deliverables becomes a Cowork team job without retyping.
When a Chat request has several files or outputs, a card under the answer shows why a team suits it, the plan and what the team will hand over, with "Run this as a team" and "Not now". Pressing it starts a Cowork task in the same workspace. The Chat then shows "Continued as a team in" with Open, and the Cowork task shows "Started from" with Open.
What stands out: The offer is made from the shape of the job, with a one-sentence reason and a draft plan of steps and deliverables.
Limits: Never offered for a single answer, one file or a quick change, and never in Code, Cowork or a private chat. Never offered twice for the same job unless you ask.
Research council (research with more than one model) #
Two to four models check the same question, so one model's miss is caught by another.
Pick models on the Research council card, press Start, then read Agreed, Disputed, "Found by one only", Models and Sources, with "Second opinion" on unless unticked.
What stands out: Agreement and dispute are shown as separate lists, not blended into one answer.
Limits: It runs on your own keys. With no other model there is no second opinion.
Extended research plan #
You set the questions and places before money is spent on a large research job.
Edit "Questions to answer", "Where to look" across The web, NeuralVault notes and Files in this project, pick Budget Standard, Larger or Largest, then "Start research".
What stands out: The budget is picked before the work starts, from named options.
Limits: Budgets are larger than Deep research at twenty searches and thirty page reads.
Wide research over a list #
A list becomes a filled spreadsheet, row by row.
Give a pasted list or a CSV or text file and what to find for each item, then watch the Research workbook fill with a Sources column and Status Done, "Not found and why", or "To do".
What stands out: Each row carries its own source and status, so unfinished rows are visible as "To do".
Limits: A cell stays empty unless a source says it. A stopped job keeps what it had.
Check sources #
Claims in an answer are checked against their pages again.
Under an answer that cites pages, or right click then "Check sources", read each claim as Verified or Unconfirmed, then pick a claim for its passage and "From page".
What stands out: Each claim is judged on its own passage, not the answer as a whole.
Limits: The guide states no limit for this feature in the lines opened.
Library #
Conversations, designs and files made or received across Chat, Cowork, Design Studio, NeuralVault and bots in one searchable list.
Open Library in the sidebar. Heading "Library", line "Find what you made, and pick up where you left off." Recent designs show first as cards with a cover picture. Under that, "All work" or the chosen kind, with a count. Filter buttons appear only for kinds present: Presentations, Websites, Documents, Posters, Images, Videos, Audio, Designs, Tasks, Files, Conversations. Search box: "Search your work". Order: "Recently edited first" or "A to Z". A design opens beside its conversation, a file in the file viewer, a conversation in Chat. The file viewer offers Download, "Open in app" and "Show in folder".
What stands out: It is not a folder browser. It joins work from five places, keeps a cover picture for designs, and reopens each item where it was made.
Limits: Capped at 600 items. A text preview shows the first 200,000 characters. A file that cannot be previewed says "No preview for this kind of file. Download it or open it in its app." Vault private notes are left out.
Agents #
Reusable agents (a saved set of instructions and tools for one kind of work) that can start new conversations instead of being explained each time.
Open Agents and skills, Agents tab: "An agent is a reusable set of instructions and tools for a kind of work. Start from one of ours and make it yours; it works in Chat and Cowork." Actions: "Use for new conversations", "Make a copy", "Open folder", "View definition" and Delete. Each agent you made is one folder. Copies, edits by hand and deletions are picked up on the next roster read, without a restart.
What stands out: An agent is a folder you own. It can be opened, copied and edited, and the change applies to the next conversation on its own.
Limits: Only agents you made can be copied or deleted. A built in agent id cannot be taken over by a folder. Delete goes to Recently deleted. This page holds no agent editor of its own.
Skills #
Short instruction sets for one kind of job, such as a weekly report style, picked up by itself when the job fits, or called by name.
Agents and skills, Skills tab: "Skills are instructions NeuralSpark follows for a kind of job, such as your weekly report or how you like quotes written. It uses one by itself when it fits, or type / and its name." Groups: Your skills, Built in, "From this project", "From Claude Code". Switch on or off, "Make your own", Import, Delete your own. Ten built in skills ship, including brand-assets, business-documents, creative-design, email-ready, frontend-design and work-package. In a conversation, saving a skill drafts a card first with Save, Edit and Discard. Nothing is written until Save.
What stands out: A skill is a plain folder with a SKILL.md file. It can be imported, switched off without deleting, and drafted from a finished conversation.
Limits: Name must be lowercase words joined by hyphens. Description fits 1,024 characters, instructions 40,000. A zip over 20 MB, over 300 files, or with a password, link or path that leaves the folder is refused whole. Project and Claude Code skills can be seen but not switched or deleted here.
Automations, the Yours tab #
A written job runs by itself on a trigger you pick, in Chat, Cowork, Code or Git, and the result goes where you chose.
Open Automations. Lead: "Work NeuralSpark does by itself: on a schedule, when a web page changes, when a file lands, when it starts, or after another job. Each run is a task you can open, and the result goes where you choose." Press "Create an automation". "Runs in": Cowork, Chat, Code or Git. "Starts" offers schedule, web page change, file lands, email arrives, NeuralSpark starts, after another job, web address called, new in a connected service, form on a live page, or "Run now". Under the form: "Only tell me if", "Send the result to" (notice, note in a folder, Slack, Teams, Discord, Tell NeuralBot, Telegram), "If a run fails", "Works in folder", "Where it runs" (this PC or an always-on helper). List filters: All, Active, Paused, Completed, Routines. History holds the last 10 runs. The Git kind commits every change in a folder and can push after, but never commits what looks like a password, key or token.
What stands out: One place holds ten trigger kinds, with per run history, retries and delivery to a notice, folder, chat apps, Telegram, NeuralBot or a NeuralVault project Updates page.
Limits: A run on an always-on helper uses the default model and its own web search, and cannot use files, apps or connected services on this PC. Closed runs need Windows and you signed in. The first check of a watched page only seeds what is there.
Ready-made automations and Teach it by showing #
Common recurring jobs from a filled form, and browser steps you showed once repeated without AI on each run.
Under "Start from a ready-made one", six show first and "Show all 11" opens the rest: Monday morning briefing, Price watch, Competitor news, Invoice inbox, Website check, Month-end report, Tender and job watch, Code health check, Security check, Weekly post drafts, Start of day. The line beside it says 'Or ask in Chat: "every Monday at 8, …"'. Choosing one fills the form. Nothing is saved until you save. In the form, "Teach it by showing" appears when the teach tool is there. A draft from Chat appears as a card with Turn it on, Test run and Edit.
What stands out: The set is written for Australian small business work, including GST in the invoice and month-end jobs. The security check reads only and names the file and line.
Limits: A template that asks for a folder stops the save until a folder is chosen. If the site changes, the form says to show it again. The gallery is fixed at eleven in this code.
Automations, the From conversations tab #
Reminders and recurring tasks set by talking in any conversation, listed in one place to change or delete.
In Automations choose "From conversations" beside Yours. Lead: 'Ask in any conversation, for example "remind me at 4 pm to send the invoice" or "every weekday at 9, check the build". The conversation keeps the task and gets a message when it is due; it is listed here to change or delete.'
What stands out: A reminder set in passing in chat is not lost in that chat.
Limits: This tab holds only conversation set tasks. Full trigger and delivery choices sit in the Yours form.
The NeuralSpark browser #
Web pages opened, read and worked in a browser beside the conversation that you can watch and take over.
Ask for a page that needs clicking or a form. The browser opens beside the conversation. A read lists what can be used by number, then the page words, sections and script errors. Actions name those numbers: click, type, choose, hover, drag, key, scroll, find, wait, fill a whole form. For a map or canvas, a screenshot gives a picture and actions use x and y. Your own browser tabs are listed by number. Tab 0 returns to NeuralSpark's own browser. Downloads and a page saved as PDF land in the task's folder, in Downloads. Settings cover allowed sites, cookie banner refusal, shared sign-ins, and search engine Google, Bing or DuckDuckGo (Google default).
What stands out: You watch the same page, can take it over for parts only you may do, and get it back in one flow. Page words are treated as information, never as instructions.
Limits: You are asked before anything is sent, booked, bought, paid, uploaded or deleted. A password, one-time code or robot check is never typed; an ask-person step waits up to 30 minutes. One task uses the browser at a time, with a 90 second hold for another. Google sign-in may be refused inside the built in browser.
Show me once #
Show a web job once and it repeats with no AI.
Press the round button beside Snip, do the task under the red Recording bar, Stop, then "Save what you showed". Set Fixed or "Changes each run", and Save. It is the same repeat path as "Teach it by showing" in automations, kept here as the browser entry.
What stands out: The repeat runs the recorded steps, so each run costs no model call for the steps themselves.
Limits: Only NeuralSpark browser pages record. Password and card boxes are never kept. A changed page hands over with "Your turn" and Done.
Saved passwords #
Saved sign-ins fill without the AI ever seeing them. See also Safety and control.
Settings, Browser, then "Saved passwords", then "Import from Chrome". On a sign-in page press Fill on the "Saved sign-in" bar.
What stands out: The password is filled by the browser layer. No copy or show action exists for it.
Limits: Nothing is offered on a page not sent securely apart from localhost. No password can be copied or shown. Chrome keeps its own after "Delete all here".
Save page to NeuralVault #
A web page becomes a readable vault note.
Press "Save this page to NeuralVault" or right click then "Save page to NeuralVault". Pick "Save in", Template and Pictures, and find a Resource with status "To read".
What stands out: Article, recipe and product templates keep their own fields, so the saved note is structured, not a text dump.
Limits: The guide lines opened state there is no Chrome or Edge extension for this path and pages save from NeuralSpark browsers. The separate browser extension in This PC is a different, paired path.
Browser tabs, history, favourites and settings #
Your own tabs, history, downloads and favourites behave like a normal browser, under your control.
"New tab" or Ctrl+Alt+T opens a tab. "Share my tabs and sign-ins" lets the AI see names and addresses. Right click a tab for "Close tab", "Duplicate tab", "Mute tab", "Save page as PDF" and "Reopen closed tab". Ctrl+F is "Find on this page". History has "Search your history". Downloads has Open and "Show in folder". The star or Ctrl+D marks a favourite. Settings, Browser holds "Show where NeuralSpark clicks", "Decline cookie banners for me", "Share my tabs and sign-ins", Search engine, "Sites the browser may use", "Close the browser", "Clear history" and "Sign out of every site".
What stands out: Sharing is off until you turn it on, and localhost always works even on "Only these sites".
Limits: Switched off, it neither sees nor uses your tabs. Tabs never share camera, microphone or location. Programs and scripts in Downloads have "Show in folder" only.
Designs: decks, documents, posters and web pages #
A presentation, document, poster or web page from a request in Chat, checked as the customer will see it, exported as PDF, editable PowerPoint or Word, pictures or a web file.
Ask for a deck, report, poster or page. It opens beside the conversation in a live panel. The start asks for kind, title, paper or format and a brief: purpose, who it is for, tone, what it must have, the one message, what the reader should do next. Writing fills slides, items, a poster or sections. A later change changes only what is named, and only changed slides or pages are drawn and checked again. Export runs the check first. Before a first export, an independent critic scores the piece against the brief, 8.5 of 10 with nothing under 8, for at most four looks. Export takes PDF, Word, PowerPoint, pictures, one HTML file or a website folder in one call.
What stands out: Built as structured parts in a style or the customer's brand, checked from rendered pages, scored by a second look, and only then exported.
Limits: A customer's own PDF that must keep its layout and photos is not rebuilt as a design. Print bleed or crop marks apply only to A4, A3 and A2 print posters. The critic can stop after four looks or two rounds without a better score, with what is still open said plainly.
House styles, brand kits, pictures and Publish #
A business look kept once in colours, fonts, logo and house rules, applied to each new piece, with a finished page put on your own web host.
A brand kit can be saved, listed, applied and removed. Applying sets colours, fonts, logo and footer. A house style holds style, colours, fonts, texture, photo treatment, banned looks and lessons from earlier jobs. Four starters ship: NB, Kaitiaki and NeuralSpark, Family and fun, The Family Gazette. Pictures can be made, or a photo treated (sepia, aged, halftone, mono, duotone), cropped, cut out or stripped of a key colour. Publish puts a website, any HTML page with its files, or a NeuralVault folder as a site onto Netlify, Cloudflare Pages, GitHub Pages or your own server over SSH. In the guide: press Publish, pick the host, Connect, then Publish, Republish, Copy, Open, "Use your own domain" or Unpublish.
What stands out: The brand is read from the customer's own document at the start, saved as a kit, and held by the check, rather than recoloured by hand on each piece.
Limits: Publish needs an account or server with the named host and its token or sign-in. Tokens stay protected and never show again. GitHub Pages is public. A password works only on Netlify paid plans. Unpublish asks first. A Family-private photo needs your OK each time.
Local office files: Word, Excel and PowerPoint #
New Word, Excel and PowerPoint files as real files, and your own files changed in place while keeping sheets, styles and figures intact.
A new document or deck is designed first, then exported as docx or pptx. A new workbook is built with the workbook kit. A file you give to change is opened with its Office library (openpyxl, python-docx or python-pptx, run with the Python that ships with NeuralSpark, so nothing is installed). Only what was asked changes. The result saves beside yours ending in " (updated)", unless you asked to change the file itself, and is reopened and compared sheet by sheet. Amounts stay as numeric or date cells with a number format. Totals are real formulas. Word changes show as tracked changes you accept or reject, unless you asked for a clean copy.
What stands out: A changed workbook is edited, checked by formula and compared with the original, rather than rebuilt from its text, which would drop charts, validations and formats.
Limits: Office libraries run inside the sandbox. A template made from a finished document must clear every old name, date and figure, named in the guide as work to check, not a promise the first pass is clean.
Talk and Dictation #
A spoken conversation in Chat, where what you say becomes a normal message and the reply is read aloud.
Start Talk in a conversation, or the Talk button or Ctrl+Shift+Space for Listening, Thinking or Speaking with Mute, "Push to talk" and End. "Share screen" or "Share camera" sends one still picture with the next spoken message, only when pressed. The mic button starts Dictation. How Talk listens and speaks is picked in Settings, Voice and Talk: this PC's speech model once downloaded, a connected OpenAI or Gemini key, or Windows voice typing and voices. Without a key, Talk runs as speech to text, the conversation, then text to speech. Audio is handled in memory, capped at 4 MB per take. Spoken replies cap at 4,000 characters.
What stands out: Talk is not a side channel. Spoken words enter the same conversation, with the same approvals and memory, and can carry one shared picture.
Limits: Live voice needs an OpenAI or Gemini key. The Grok voice entry is marked not ready and is not offered. Code, tables, cards and addresses are not read aloud. A draft in the box is never sent by Talk. Screen and camera stop straight away, with nothing kept. Speech goes only where "Listen with" says and is never saved.
Spark
NeuralBot
NeuralVault
NeuralVault comes with it
