The full stack, in plain language.
Most AI products are a chat box with a company's name on it. This is a platform: one governed, metered assistant that reaches you wherever you work, remembers what you told it, reads the systems your business actually runs on, and — when you ask it to — builds and maintains real things on the open web.
Everything below is built and running today. Nothing on this page is a roadmap, a beta, or a thing we could build for you if you asked — this list is generated from the platform itself, so it can only describe what is actually there. Where something is built and deployed but has not yet been proven in real use, its entry says so.
One assistant, one conversation, reached from a browser, a Telegram DM, a text message, an embedded widget on your own site, or an email address of its own. Same governed turn, same memory, same billing — the channel is just the door.
Hands a long job to your assistant to work on in the background. You ask for it once and carry on; the work continues after the conversation moves on.
Work that takes minutes doesn't hold up your conversation, and you don't have to sit and wait for it to finish.
For example Ask for a full revenue briefing on last quarter and keep chatting while it is put together.
Checks on a background job you already asked for — whether it is running, finished or failed, and what it produced.
You always know where a long-running request stands, instead of wondering whether it was ever started.
Opens up the other ways to reach your assistant — the live streaming web client, Telegram and email — beyond the basic question page.
You can ask from wherever you already are, and answers stream back as they are written instead of arriving all at once.
For example Message your assistant on Telegram from your phone and get the same answers you would get at your desk.
The place you ask questions and get answers. What you said earlier is carried forward, so you can follow up without repeating yourself.
You get answers that build on what you already said, instead of starting from nothing every time.
For example Ask "what did we decide about the refund policy?" and then just "why?" — the follow-up still knows what you meant.
Watches for the topics you have flagged as risky or easy to underestimate, and steers the answer into a fuller, more careful one when they come up.
The questions where a glib answer costs you most get the most thorough treatment, without anyone having to remember to ask for it.
For example Flag "termination", "refund" and "liability" so those questions always come back with the caveats rather than a confident one-liner.
Works automatically, on every relevant answer.
Built and deployed — not yet proven in real use.
Puts a decision question to two different AI models independently, then has a third model from a different provider score both answers against a fixed rubric and pick one — or say honestly that both are defensible.
On a decision that matters, one model's answer is one model's blind spots. Two independent answers plus a neutral judge surface the disagreement instead of hiding it — and when the tradeoff is genuinely unresolved, you are told that rather than handed a false winner.
For example Ask whether to prioritize marketing spend or cost cuts for the next 90 days, with your revenue, cash and pipeline figures as the supporting data.
The assistant uses this when your question calls for it.
Creates an image from a description you type, using an image model of your choice, and stores it in your organization so you can use it again.
You get artwork for a proposal, a page or a document without leaving the assistant or licensing a stock photo, and it stays in your organization for reuse.
For example Ask "generate a wide banner image of a quiet lakeside cabin at dawn, illustrated style" and use the image it returns.
The assistant uses this when your question calls for it.
Your website's chat assistant can offer interested visitors a small contact card — name, email, phone — at the moment they show intent, and each submission becomes a lead you are emailed about and can review.
Visitors who chat are your warmest prospects; this turns a good conversation into a name and an email instead of an anonymous session that ends.
For example A visitor asks about pricing for their nonprofit; the assistant answers and offers the card; two minutes later the lead is in your inbox with the conversation context.
Turns unstructured text into structured data that matches a format you defined in advance. You register the shape you need once; every extraction after that either matches it exactly or tells you plainly what was missing.
The hard part of importing messy material is not writing it out — it is getting something an import can trust every single time. Output is checked against your format before you ever see it, so a bad result becomes an honest message instead of a broken import you discover later.
For example Register the format your lesson importer expects, then paste in a page of rough lesson notes and get back a file the importer accepts as-is.
The assistant uses this when your question calls for it.
Separate spaces for separate jobs. Each workspace keeps its own conversations and its own set of turned-on features.
Work on one project doesn't bleed into another, and you change context in a click instead of scrolling past unrelated history.
For example Keep one workspace for marketing and another for support, so the two never mix.
Every conversation is saved for you automatically. You can rename one, hide it, or come back to it days later.
Nothing you asked is lost when you close the tab, and a conversation you care about is easy to find again by name.
For example Rename a long thread to "Q3 pricing review" so you can find it again next week.
Three distinct kinds of memory, deliberately separate: the verbatim transcript of what was actually said, the durable insights distilled from it, and the confirmed facts about your business. You can read every one of them, and delete any of them.
Lets your assistant write down a durable fact about how the business is doing, as a proposal. Nothing it writes counts until a person confirms it.
Useful observations get captured while they are fresh, without an unreviewed guess quietly becoming something the business treats as true.
For example After a look at last quarter it proposes "renewals are carrying growth more than new customers" — noted, not yet trusted.
Brings confirmed business observations into answers — but only the ones relevant to what you actually asked.
You get answers that reason from what the business has established, without every saved note being dragged into every reply.
For example Ask "should we push renewals this quarter?" and the confirmed observation about renewals carrying growth is already in view.
Works automatically, on every relevant answer.
Confirms or rejects the business observations waiting for review. Only confirmed ones are ever used in answers.
A person decides what the business treats as true, so answers rest on reviewed judgement rather than an unchecked inference.
Finds the parts of your earlier conversations that relate to what you are asking now, and puts those exact words back in front of the assistant.
Long conversations stop losing their beginning, and picking a thread back up days later does not mean re-explaining what was already decided. When a remembered line is only part of the story, the assistant can open that whole earlier conversation and read what was said around it.
For example Three weeks after you settled on a launch date in a long chat, "remind me what we agreed about the launch" quotes what you actually said then.
Works automatically, on every relevant answer.
Short notes kept about things worth remembering — a preference, a decision, a recurring fact. You can read them, edit them, or delete any of them.
You stop re-explaining the same background every time, and you stay in control of what is remembered.
For example Once it's noted that your financial year ends in March, "how did last year go?" means the right year without you saying so.
Saves something from a conversation as a durable note that can be used later.
The things you would otherwise have to repeat get written down once, deliberately, by you.
For example Say "remember our support hours are 8 to 6 Eastern" and later answers use it.
Brings the notes that were saved into the answers where they are relevant, so context carries forward between conversations.
A note only earns its keep if it turns up at the right moment. This is the half that does the turning up, without you having to remember the note exists or go looking for it — and it can search its saved memory on its own when a question calls for it, rather than only using what happened to come up.
For example You mentioned months ago that the Tuesday crew handles commercial jobs. Ask "who should take the Riverside contract?" and that comes back with the answer, unprompted.
Works automatically, on every relevant answer.
After a substantial exchange, what was worth remembering is noticed and saved without you having to ask. It reads the recent conversation rather than only your last message, so something you build up across several messages is kept as one complete sentence.
Useful context gets kept even when you don't think to save it, and even when you never said it all in one go. The same statement is never saved twice, so your memory page stays worth reading, and you still get the final say on what stays.
Ask about revenue, traffic, ad spend or your pipeline in plain language and get an answer computed from your own data. We read the totals and throw the rest away — the raw records never land on our disks. And where a service of yours publishes its own tools, you can connect it directly and let the assistant use them.
Brings the tools from other services' MCP servers — the ones your organization registers and trusts — into your assistant, so it can use them in a normal conversation.
Your assistant can act in the other services you already use without you leaving the conversation or Phasma having to build a connector for each one.
For example Register your issue tracker's MCP server, then ask "what's assigned to me this week?" and get the answer from the tracker itself.
The assistant uses this when your question calls for it.
Answers questions about your Google Ads campaigns — spend, clicks, conversions and cost per click — from your own account's numbers.
You ask a plain question and get your real campaign numbers back, instead of exporting a report and reading it yourself.
For example Ask "what did we spend on ads last month, and what did it get us?"
Works automatically, on every relevant answer.
Answers questions about your website traffic — visits, visitors, and where they came from — using your connected Analytics property.
Traffic questions get answered in the same conversation as everything else, so you can ask what changed without switching tools.
For example Ask "where did last month's traffic come from, and how does that compare with the month before?"
Works automatically, on every relevant answer.
Built and deployed — not yet proven on a real site.
Answers questions about how your site does in Google search — clicks, impressions, position, which pages and searches bring people in, and whether your sitemaps and key pages are being read — from your own Search Console property.
You get your real search numbers in the same conversation as everything else, with the current week clearly marked as partial and the date the data runs through stated.
For example Ask "which searches brought people to the contact page last month?"
Works automatically, on every relevant answer.
Answers questions about your HubSpot pipeline — new leads by source and stage, new contacts, deals closed, and sales-email activity.
You can ask how the pipeline is moving without building a report, and the answer never exposes anyone's contact details.
For example Ask "how many leads came in last week, and where did they come from?"
Works automatically, on every relevant answer.
Answers questions about your Facebook and Instagram ad campaigns — spend, clicks, conversions and cost per click — from your own account's numbers.
You can compare how the paid channels are doing without opening a separate dashboard for each one.
For example Ask "which did better last month, our Meta ads or our Google ads?"
Works automatically, on every relevant answer.
Lets your assistant read files and folder structure from the code repositories you have connected. Read-only — it never writes or changes anything.
You get answers grounded in what the code actually says, instead of what someone remembers it saying.
For example Ask "what does our deploy script actually do?" and the answer quotes the file.
The assistant uses this when your question calls for it.
Answers questions about your revenue — recurring revenue, growth, churn and more — from your connected Stripe account.
You can ask "how are we doing?" and get real numbers back, in the same place you ask everything else.
For example Ask "how is this month tracking against last?"
Works automatically, on every relevant answer.
Built and deployed — not yet proven in real use.
Answers revenue questions from a sales file you uploaded, for when the numbers don't live in a system anyone can connect to.
You get the same revenue answers with no integration at all — a spreadsheet export is enough to start.
For example Upload last year's sales export and ask "which month was strongest?"
Works automatically, on every relevant answer.
Upload what you have, ask questions of it, and have the assistant compose, revise and typeset new documents that draw on it.
Where finished documents live — the ones written for you and the files kept alongside them. Each one keeps its earlier versions, so you can see what changed.
Finished work has one home instead of being buried in a conversation, and you can always show which version someone actually received.
For example Open last month's proposal, compare it with the current wording, and download the one you sent.
Turns a conversation into a real document — drafted, revised, rendered to a finished PDF, every version kept, and an upload can become one. It also designs WEBSITES: ask for one and the assistant composes a live mockup from the site builder's real blocks — preview it, tune its colors, sizes and fonts, then have it built as a real site.
The work of turning a decision into a document someone can actually be sent — the formatting, the layout, the export — stops being a separate job you do afterwards.
For example Ask for "a proposal for the Rivera job, as a PDF" and the finished file comes back in the reply.
The assistant uses this when your question calls for it.
Lets you create knowledge bases and add or remove their documents.
You keep the reference material current yourself, so answers stop quoting last year's price list.
For example Upload this year's service price list and delete last year's — answers switch over from then on.
Lets you see which knowledge bases exist and what is in them.
You can confirm the material an answer should be drawing on is really there, before wondering why it isn't.
Pulls the relevant parts of your knowledge bases into an answer at the moment you ask. This is the piece that actually makes answers draw on your own material.
Answers quote your hours, your prices, your policies. When the first passages are not enough, the assistant searches again in its own words or reads the whole document a passage came from. Without this, a fully loaded knowledge base is never consulted, no matter how much is in it.
For example Upload your service price list, then ask "what do we charge for a deep clean?" — the answer comes from the document.
Works automatically, on every relevant answer.
Prepares the documents you add to a knowledge base so they can be found later — reading them, splitting them up, and storing them for search.
Your own material becomes searchable. Storing it is only half the job, though — actually drawing on it in answers is what Knowledge-base answers does.
The assistant uses this when your question calls for it.
Reads back the full text of a document already in view — one you uploaded, one remembered from earlier, or one the assistant composed for you (by its number).
You can ask detailed questions about a long file without pasting it in, and the answer comes from the real text rather than a summary of it.
The assistant uses this when your question calls for it.
Built and deployed — not yet proven in real use.
Lets your assistant open a public web page you name — or one it needs to check — and read it as plain text, the way you would in a browser.
You get answers from what the page says today rather than from what the assistant remembers, and you can see which pages it read under the answer.
For example Paste a supplier's terms page and ask "what is their cancellation notice period?" — the answer quotes the page.
The assistant uses this when your question calls for it.
Attach a file to a question and ask about it directly. Text files, spreadsheets, PDFs and images all go in the same way.
You ask about the actual file instead of describing it, with no copying and pasting of long extracts.
For example Attach a supplier invoice and ask "what's the payment deadline on this?"
This is the part nobody else does. The same assistant can generate a complete WordPress site, migrate an existing one, audit it for privacy compliance, fix what it finds, and then keep operating it — every change proposed for a human to approve before it happens.
Drafts new posts and pages for your connected WordPress site, or proposes careful edits that change only the words you asked about. Every proposal waits for a person to approve it, and whoever asked is told when it goes live, is declined or is undone.
You get drafts written for the real site, in the right shape, without anything gaining the ability to publish behind your back — and nobody has to chase what happened to a request.
For example Ask for "a services page for our new emergency callout offering" and review the draft before it goes anywhere near live.
The assistant uses this when your question calls for it.
Scans what a website does in public — cookies and trackers set before anyone consents, whether consent controls work, whether the privacy policy is reachable — and reports what it saw, never a compliance verdict. Rescan to see exactly what changed, or turn a finished scan into a draft privacy policy staged for approval.
You see what a regulator or a careful customer could see from the outside — on your own site or a prospect's — with nothing to install, and get a policy draft that matches what was actually found instead of starting from a blank page.
For example Ask "scan example.com and tell me what it sets before consent" — then, once it's done, "draft a privacy policy from that scan".
The assistant uses this when your question calls for it.
Lets your assistant look at your connected WordPress site — its pages and posts, its plugins and theme, and its safe configuration. Read-only.
You get answers about how the site is actually set up, rather than what someone thinks was installed two years ago.
For example Ask "what plugins are running on the site, and are any of them switched off?"
The assistant uses this when your question calls for it.
Builds a website through conversation. Say what you want and the assistant works out whether it is a brochure site, a migration, a larger custom build, or a rebuild of a WordPress site from its backup (content moved into the page editor, the site put on a standard theme), then asks for what it needs and stages the result.
The whole build happens where you are already talking, and the site appears at a private staging address you can look at before anyone else sees it. Going live is still a deliberate, separate act by a person.
The assistant uses this when your question calls for it.
Reads a business's public website and builds a reusable profile of what it publishes — who they are, how to reach them, what they offer, who is on the team, and how the site looks and performs.
You get a structured picture of a business in one step — useful for researching a prospect, seeding a knowledge base, or building a demo site that matches the real one.
For example Ask "build me a profile of riveraplumbing.com" before a first sales call.
The assistant uses this when your question calls for it.
Creates and manages files inside your WordPress site and drives a curated set of site operations — switching a theme, turning a plugin on or off, installing a plugin package your administrator has staged. Changes wait for a person to approve them unless you have deliberately turned that requirement off.
Real changes to the site can be built and staged for you, while approval stays the point where a person decides. Be deliberate about who gets this one: writing a file into a site is as powerful as holding the keys to it.
The assistant uses this when your question calls for it.
Every turn is metered, attributable and auditable. Coaches see only their own clients, each business sees only its own data, and there is a record of what was done on your behalf.
Applies your house tone and formatting to every answer — how formal to be, how long to run, whether to use bullets.
Answers come out sounding like your own people wrote them, so you can paste one into an email without rewriting it first.
For example Set "short, plain English, no bullet lists" and a three-paragraph answer comes back as three tight sentences.
Works automatically, on every relevant answer.
What you've used and what it cost. Your balance, your recent activity, and the record of what was charged for what.
You can see where the spending actually went before the bill arrives, instead of guessing from a single total.
For example Find the week the usage jumped, and see what was being asked at the time.
Puts your own standing instructions in front of every question — who the assistant is for, what it should assume, what it must never do.
You stop repeating the same background in every conversation, and everyone gets answers shaped the same way.
For example Say "we are a dental clinic in Montreal; always answer in the client's language and never give medical advice" and every answer respects it.
Works automatically, on every relevant answer.
Lets you choose which model answers your questions, instead of always using the standard one.
You can trade cost against depth deliberately — a stronger model where it matters, a cheaper one where it doesn't.
Works automatically, on every relevant answer.
Keeps a reviewable record of each question and answer, with timing and size details attached.
You can answer "who asked what, and when" months later — for a review, a client question, or a compliance request — without relying on anyone's memory.
Runs after each answer, in the background.
Sends a signed notification to a web address of yours whenever certain things happen, so another system can react to them.
Your other tools find out immediately instead of on a schedule, and the signature lets them prove the message really came from here.
Sends results onward to where you need them.
Further capabilities available in the stack.
How to pay: save a credit card, buy credits, and turn on automatic top-up so your balance never runs out.
Your assistants keep answering without stopping for a manual top-up, and you decide the monthly ceiling.
For example Save a card, buy a starter pack, then turn on auto top-up with a monthly cap so the balance refills by itself.
Puts several of your assistants in one of your own Slack channels. Write "@Phasma <name>: your request" to talk to one, and an assistant can ask another for help by posting its request in the thread, where everyone sees every ask and every answer.
Work that needs two specialists happens in one conversation your team can watch. Each thread has limits (by default 8 assistant turns and 100 Ecto); reaching one posts "Paused — … Mention one of us to continue." and your next message starts fresh.
For example In #website, ask "@Phasma wp helper: add a hero image to the About page". The WordPress assistant asks the art director assistant for the picture, the image appears in the thread, and the WordPress assistant proposes the change for your approval.
The assistant uses this when your question calls for it.
Lets your assistant open and read web pages in a Chrome window on your own computer, and click or type in them, asking you first unless you turned on full autonomy, from the phasma-code terminal you paired.
The assistant works from the page you are signed into, in a window you watch, instead of asking you to copy things out of it.
For example Ask "open the Stripe dashboard and summarise this month's failed payments".
The assistant uses this when your question calls for it.
Lets your assistant change files in a project on your own computer, from the phasma-code terminal you paired with your seat. You see the change and approve it first.
The assistant can make the edit it proposes, with your version control as the safety net.
For example Ask "add a --verbose flag to the CLI and update the help text".
The assistant uses this when your question calls for it.
Built and deployed — not yet proven with a real coding task.
Lets your assistant open a coding task — a brief, a repository and a budget — that your paired phasma-code workstation picks up and works on without anyone at the keyboard.
A coding job can run while nobody is at the terminal, inside the workstation's own rules, and stops at its budget. A person reviews the draft pull request it opens.
For example Ask "open a coding task on 42kohd/phasmas-platform: reconcile the stale followups line".
The assistant uses this when your question calls for it.
Once in a while the assistant asks you one short question about something it worked out on its own but never heard from you, so it can stop guessing.
A wrong guess is caught before it shapes more answers, and a right one becomes something you said rather than something it assumed.
For example "I put two things together and concluded you run errands on Fridays. Is that right?" — you reply "Not quite, only the first Friday of the month", and that is what it remembers.
Works automatically, on every relevant answer.
It never makes up facts about your business. If it doesn't have something on record, it says so and asks, instead of guessing a price, a name, a date or a policy.
You can put the assistant in front of staff and clients without worrying that it will invent a fee, a deadline or a promise on your behalf. What it does say about your business comes from your own documents, your notes, or what you told it.
For example Ask "what does the deluxe plan cost?" when no price list is on record and it answers "I don't have that on record, so I won't guess", lists what it does have, and offers to remember the price once you tell it.
Works automatically, on every relevant answer.
Built and deployed — not yet proven in real use.
Sends the leads you track in Phasma into your organization's HubSpot as contacts. Sending the same lead again updates its contact instead of creating a duplicate.
HubSpot stays the place your contacts are kept, and nobody re-types a lead that was already caught and qualified here. A field you filled in by hand in HubSpot is never blanked by a later send.
Lets a member of your team answer website visitors who asked for a person — a queue of waiting conversations, the transcript so far, the page the visitor is on, and a suggested reply drafted from your own content that they edit before sending.
Your people answer only the conversations that need them, with the context already gathered and a draft to start from, so a visitor gets a real answer in minutes.
For example A staff member sees a waiting visitor asking about delivery to their town, presses Suggest, corrects one detail in the draft and sends it, then hands the chat back to the assistant.
Works automatically, on every relevant answer.
Your website's chat assistant can hand a visitor to a real person on your team when they ask for one or when it cannot help. If nobody is online, it offers to have someone contact them and takes an e-mail address or phone number instead.
The assistant answers the everyday questions from your own content, and the conversations that need a person reach one with the whole chat already in front of them, instead of the visitor starting over by e-mail.
For example A visitor asks whether you can take on an urgent project next week. The assistant checks who is available; a colleague who is online picks it up, sees the page the visitor is on and everything said so far, and answers in the same window.
Works automatically, on every relevant answer.
Built and deployed — not yet proven with a real radio.
Lets your organization connect a Meshtastic bridge, so a radio can message your assistant with no internet at all.
Your assistant stays reachable off the grid — at an event, in the field, or in an outage.
For example From a hiking trail with no cell signal, a radio asks for the weather window and gets it.
A help desk your staff can ask about everyday computer, printer, Wi-Fi, e-mail and meeting problems. When it cannot fix something, it passes the problem to your IT contact with what was tried and what was seen.
The small daily questions get answered on the spot, and the ones that need a person arrive as a finished diagnosis with a name and a way to reach the asker, instead of "my computer is broken".
For example A staff member's laptop asks for a recovery key after an update. The help desk stops, takes their name and extension, and your IT contact gets one e-mail saying exactly which laptop, what appeared on screen and how to reach them.
Works automatically, on every relevant answer.
Talk to your assistant instead of typing, and hear its answers read aloud — on the web, on your phone, and on any device paired to your organization.
You can ask while your hands are busy or you are away from a keyboard, and long answers are read to you as they are written instead of after the whole reply is done.
For example Ask "what's on my plate this afternoon?" while you walk to a meeting, and listen to the answer in your earbuds.
Puts the documents you pin in front of the assistant on every single answer, in full, marked as the reference it must follow.
Prices, terms and links are never "not found" because a message was too short to search on — "yes" gets the same facts as a long question.
For example Pin your price list and trial terms; a lead who texts "ok" gets a reply that quotes the current price, not an invented one.
Works automatically, on every relevant answer.
Lets your assistant read, list and search files in a project on your own computer, from the phasma-code terminal you paired with your seat.
The assistant answers from the code as it actually is on your machine, not from memory.
For example Ask "summarise what the README says about deploys".
The assistant uses this when your question calls for it.
Lets your assistant run commands — tests, builds, git — in a project on your own computer, from the phasma-code terminal you paired with your seat. You approve each one first.
The assistant can run the test suite and react to the real failure instead of guessing.
For example Ask "run the unit tests and fix whatever fails".
The assistant uses this when your question calls for it.
Built and deployed — not yet proven in real use.
Connects your organization's Salesforce account to Phasma, so the leads you track here can be sent into Salesforce as Lead records.
Your Salesforce stays the place leads are kept, and nobody has to re-type a lead that was already caught and qualified here.
Reads documents that defeat automatic text extraction — scanned directories, fire-insurance plans, census pages, archival maps — by looking at each page the way a person does. It writes the result out as data files you can download.
Old and archival material is usually a picture of a page rather than text. Automatic text recognition either fails on it or, worse, returns confident nonsense. This looks at the page instead, flags what it cannot read rather than guessing, and checks every file it writes against the shape you asked for, so nothing malformed reaches whatever you feed it into.
For example Upload thirty scanned pages of a 1990 street directory and get back one data file per page, plus a list of every entry that could not be read cleanly.
The assistant uses this when your question calls for it.
Built and deployed — not yet proven in real use.
Hands a task from one assistant seat to another — a specialist your organization set up — and brings the specialist's answer back into the conversation you asked from.
The specialist works with its own instructions, tools and memory, and remembers what it learned for next time, while you stay in the conversation you started.
For example "Hand this to the coding seat: check whether the release script still stamps UTC dates" — the coding seat takes it on, and its answer arrives in your thread.
The assistant uses this when your question calls for it.
Lets your organization connect a phone number of its own, so people can text your assistant and get answers back by text.
People who never open a browser can still ask and be answered from their phone.
For example A volunteer texts "what time is the food bank open Saturday?" and gets the answer by text.
Creates a short video clip, four to eight seconds long, from a description you type or by adding gentle motion to a photo you attach, and keeps it in your organization's Videos page where you can watch, download or delete it.
You get a hero background, a living version of a product photo or an animated logo without hiring a videographer or leaving the assistant. The first time you ask, it tells you what a second of video costs before it makes anything.
For example Attach a photo of your storefront and ask "animate this with slow, gentle motion, about six seconds, wide" — then download the clip for your homepage.
The assistant uses this when your question calls for it.
Built and deployed — not yet proven in real use.
The assistant can search the web in the middle of answering you, read what it finds, and tell you which pages it used.
Answers about things that change are checked against the web instead of the model's memory, and you can open the sources yourself.
For example Ask "what's the latest version of WordPress?" and it searches, answers with the version number, and lists the page it took it from.
An example engagement. Every step below is something the platform does today.
A plumbing company with eight vans wants a new website. The old one was built in 2016 by someone who has stopped answering emails.
Monday. The owner describes the business in a chat — what they do, which towns they cover, what they want people to call about. No forms, no design brief. By the afternoon there is a complete website to look at: real pages, real structure, written in their own words rather than filler.
Tuesday. They ask for the emergency callout service to be more prominent. The change is drafted and shown to them before anything goes live. They approve it. It goes live.
Wednesday. They connect their payment system. Now "how did last month compare to the same month last year?" gets an answer from their own numbers, in a sentence, without opening a dashboard or exporting anything.
Thursday. A customer enquiry arrives through the new site. It is captured, the owner is emailed, and the details are waiting for them — no plugin bought, nothing configured.
Friday. They ask whether the site is doing anything it shouldn't with visitors' privacy. It gets checked from the outside, the way a regulator would see it, and they get a plain list of what was found and what to do about it.
The following month. Half-remembering a conversation about callout pricing, the owner asks how they had settled it. Back comes what they actually said, weeks earlier, quoted, with the date on it. Nobody wrote it down. It simply remembered.
One assistant. One conversation. No agency retainer, no five separate tools, and nothing happening to their business without them seeing it first.
None of this is sold as a take-it-or-leave-it plan. You pick the parts you actually want, you pay for those, and you add more when you need them — nothing switches itself on behind you, and nothing is bundled in to pad an invoice.
Not sure which parts you need? That is usually a ten-minute conversation, not a sales process. Email jer@phasma.ca and describe the business — or open an account and start with one piece.