Ten coworkers. Ten toolboxes. One memory.

Most AI assistants are one model with one giant toolbox and a system prompt asking it nicely to behave. Unyo is ten named specialists, and the thing that separates them isn't personality — it's access. Ashley cannot open your inbox. Not "shouldn't". Cannot. The email tools are never loaded into her session, so there is no instruction you could write, and no prompt injection anyone could plant, that would hand her a message.

That boundary is the product. Everything below is what each specialist owns on the other side of it.

The team

Ashley — Social media

Your channels stay alive without you living in them.

• Ask for a post and get an editable draft card: caption, hashtags, suggested platforms and a generated image — not a wall of chat text you have to copy out.
• Ask "when should I post?" and get an hour and a day per platform, ranked from your own analytics cache. Never an invented industry benchmark.
• Schedule, reschedule or cancel in plain language — "move it to Friday 10am" — and research hashtags for your niche, ranked by what has actually worked on your account.

She has no inbox, no image editing suite of her own, and no access to your revenue. → Ashley's page

Maya — Email & calendar

The only agent with your inbox.

• Read, draft, reply and batch-reply. Every reply resolves its real recipient, thread and message ID from your inbox cache on the server — Maya writes the body, the system supplies the target, so she cannot reply to an address she invented.
• Schedule a send for a future local time, then reschedule or cancel it. Your local time is converted for you.
• Attach files from Google Drive, your Neural Core or your image library — up to ten staged at once, folded into the next draft automatically.

She drafts. She does not send. Sending is gated behind a confirmation card you click. There is no auto-send setting, in any plan, on any schedule. → Maya's page

Alex — Sales

Turns your store data into outreach you can actually send.

• Pull a store summary from Shopify — orders, revenue and average order value over the window you ask for, flagged when the sample was truncated rather than quietly rounded.
• Look up orders, customers and products to answer a prospect's question with a real number instead of a guess.
• Produce product one-pagers and proposals as editable document cards, or a prospect list as a real multi-tab spreadsheet.

Shopify access is read-only by grant — all seven scopes are read_*. Alex can see your orders. He cannot change one. → Alex's page

Riley — Product

The one who writes the thing nobody wants to write.

• Product briefs, specs and release notes as editable documents with real structure — headings, tables, sections you can rewrite one at a time without regenerating the whole thing.
• Roadmap and backlog spreadsheets — multi-tab .xlsx you can open, edit in Google Sheets or export.
• Notion pages and Slack posts from a conversation, so the decision lands where your team already looks.

No inbox, no social publishing, no image generation. Riley writes and organises. → Riley's page

Tyron — Images & visuals

The only agent with a full image toolbox.

• Generate an image from a description in the aspect ratio you need, then restyle it in place — "make it warmer" — without starting over.
• Upscale to HD or cut the background out of the image you're already looking at.
• Pixel-exact compositing — logo placement, safe-area text — done as real code in an isolated sandbox, not eyeballed by a model.

The image tools are mounted for Tyron and denied to the other nine. Ashley can generate the image inside a post draft, but restyling, upscaling and background removal are Tyron's alone — ask her for those and she hands the job over rather than pretending. → Tyron's page

Sam — Customer support

Answers "where is my order?" with the tracking number, not an apology.

• Look up any Shopify order by number — "where is #1001?" — and get line items, fulfilment status, shipping address and tracking.
• Read an attached PDF, DOCX, XLSX or CSV a customer sent and answer from what's actually in it.
• Draft macros, help articles and response templates as editable documents you can reuse.

Sam sees your store. Sam does not see your inbox — that is Maya's, and only Maya's. → Sam's page

Lucy — Content writing

Research first, then write — with the sources still attached.

• Search, read the top sources, then deliver a full document — blog article, SEO article, press release, product description or newsletter — as an editable card.
• Real inline [n] citations with a references appendix, localised in ten languages. They survive editing and reloading, because they're structure, not decoration.
• Edit one section — rewrite, expand, shorten, replace, insert after or delete — without touching the rest of the piece.

→ Lucy's page

Blake — Data

He doesn't estimate arithmetic. He runs it.

• Real Python (pandas, numpy, matplotlib) over your data in an isolated sandbox — aggregations, percentiles, statistics, charts — handed back as downloadable files, not a paragraph of approximations.
• Your Stripe revenue, read-only: revenue summary by currency, available balance, recent charges, payout history. There is no charge, refund or payout tool anywhere in Unyo's code — the capability doesn't exist to be misused.
• Coder mode: ask for a script and he writes it, runs it, reads the traceback, fixes it and iterates until it runs clean — then hands you working code plus its output.

The sandbox has no internet and no environment variables. Nothing gets out of it, and nothing gets in. → Blake's page · How we secure it

Ema — Personal development

The only agent who can see everything you have pending.

• "What's on my plate?" — one cross-agent snapshot: unread notifications, posts scheduled or failed, emails waiting to go out, upcoming automations. Nobody else can assemble this view.
• Morning briefs and planning documents grounded in what your other agents actually did.
• Recurring routines set up in plain English, prepared as drafts for you to review.

→ Ema's page

Steve — Websites

Describe a site. Watch it get written. Publish it.

• The code streams live into a builder panel while the model is still generating it — you see the file being typed, not a spinner.
• Click any element in the preview to point at it, then say what you want changed in plain language. The inspect script lives in the served preview only — never in your ZIP, never on your published site.
• Publish to a real public URL at <slug>.unyo.site with an automatic TLS certificate. Edit and say "republish" and the same URL updates. Say "unpublish" and it's gone.

→ Steve's page

Why ten specialists beat one chatbot

The boundary is a runtime control, not an instruction

Here is the part that matters, and the part almost nobody else does.

When you open a conversation, Unyo builds that agent's toolbox for that request. Providers are mounted conditionally by agent id: the Stripe client is only ever spawned for Blake, the website client only for Steve, the image client only for Tyron, and the email client only for Maya. If an agent isn't the one, the provider is never created — its tools never enter the tool list, and its JSON schema never reaches the model at all.

On top of that sits a per-agent denylist that strips anything left over.

So when we say Ashley can't touch your inbox, we don't mean she's been told not to. We mean the email tools do not exist inside her session. There is nothing to call, nothing to jailbreak, and nothing for a malicious instruction hidden in a web page to aim at. Gmail and Outlook access is scoped to exactly one agent, enforced at the prompt layer and again at the runtime layer.

A single chatbot with every tool mounted has no version of this. Its only defence is that it was asked politely.

The enforced limits, listed plainly

Steve is the most restricted agent on the platform: no calendar, no Drive, no documents, no image tools, and read-only access to your Neural Core. He builds websites. That's the job.

Second-order benefits

• Smaller blast radius. A capability that isn't loaded can't be misused — by the model, by you on a bad day, or by text hidden in a page an agent was asked to read.
• Sharper work. An agent whose toolbox matches its job stops trying to solve everything with the wrong instrument.
• Honest failure. When an agent lacks a capability, the system knows — it computes what the agent is actually missing from its live tool list and tells it to hand off, instead of letting it improvise.

Delegation: how the team hands work over

Ask Ashley to analyse your revenue and she doesn't fake it, and she doesn't get given Stripe. She hands the task to Blake.

What actually happens: Blake opens his own conversation, runs the full pipeline with your tools and your credits, and you get a red-badge notification the moment the hand-off happens — not when it finishes. You can open his conversation and watch. Ashley tells you Blake is on it, and is explicitly forbidden from claiming she did the work herself.

Three things worth knowing, because they're the honest shape of it:

1. It's a hand-off, not a sub-agent call. The delegating agent doesn't receive Blake's answer back and stitch it into one reply. Blake's work lives in Blake's conversation. This is a team passing a ticket, not a model calling itself.
2. Depth is capped at one. A conversation that was itself delegated cannot delegate again. No runaway loops, ever.
3. Steve can't be handed a website. Delegation targets the nine non-Steve agents. To build a site you open a Steve conversation yourself — he can delegate out, but nobody can delegate in.

What every agent shares

• The same memory. All ten read your Neural Core — brand, products, contacts, knowledge — plus your long-term preferences and a digest of what happened across all your conversations in the last 18 hours. Architecturally shared, not shared by convention.
• Visible work. Most agents declare a plan up front and announce each step as they start it, flipping from pending to running to done live in the transcript. Expand a finished step and read the real caption or subject line it produced. (Tyron and Steve skip the checkbox plan — Tyron chains image work, Steve streams code.)
• Documents and spreadsheets. Nine of ten write real editable documents; all ten can build a multi-tab .xlsx you can download, open in Google Sheets, or export to CSV or PDF.
• File reading. Attach a PDF, DOCX, XLSX or CSV and any agent can read it. Scanned PDFs work too — OCR takes over when there's no text layer.
• 11 languages. Set your agent language once and every agent replies in it, whatever language you type in. Or just ask mid-conversation to switch.
• Honest gaps. Every agent knows which of your accounts are connected and shows you a Connect button instead of failing with an error wall.
• Refunds on failure. A tool that fails doesn't cost you credits — every paid tool refunds automatically, and an hourly janitor sweeps anything stranded.

Try it

You get 2 500 free credits at signup, no card. Reading, searching, drafting and planning cost zero credits — 73 of the 105 priced tools are free, and you're billed on real cost rather than per message. See pricing.

Some prompts to start with:

FAQ