Everything Unyo does

Unyo is a team of ten AI coworkers that share one memory of your business, hold real keys to your real tools, and hand you finished work - drafts, documents, spreadsheets, images, websites - instead of a wall of chat text.

This page is the map. Each pillar links to the page that goes deep.

The team - 10 agents, 10 different sets of keys

These aren't ten prompt presets wearing different hats. Each agent's toolbox is an access boundary enforced in the runtime, not a polite instruction in a prompt. Ashley genuinely cannot read your inbox - the email tools are never loaded into her session, so there is nothing for her to call. Only Maya has them. The same is true down the list: only Blake sees Stripe, only Tyron and Blake get the Python sandbox, only Steve can touch a website.

Ask an agent for something outside its job and it hands the task to the right coworker - who opens its own conversation, does the work, and pings you when it's done.

Meet the team →

The Neural Core - one memory, ten agents

This is the thing nobody else does the way we do it.

Every agent reads the same business memory: your brand, your products, your contacts, your knowledge, your rules. Not "we sync it between assistants" - the code that builds the memory view literally ignores which agent is asking. Every agent gets a byte-identical picture.

Tell Maya "never use emojis with clients" and Ashley knows it on her next run. The long-term memory is queried by you, not by agent. There's a shared [TODAY] digest too, spanning every conversation and every agent for the last 18 hours - so no agent starts its turn blind to what the others just did.

• Deleted stays deleted. Remove a memory and a tombstone tells the extractor "do not re-extract this."
• It won't learn the AI's own inventions. The extractor is forbidden from mining assistant messages - it only sees an 80-character topic hint of them.
• Your memory can't be forged. Authenticated users get read and delete permissions on their memory. There is no insert or update. Only the internal extractor writes.
• Memory is free. Every Neural Core tool - reading, searching, remembering, updating and deleting - costs zero credits.

How the Neural Core works →

Integrations - 16 live, and we asked for less than we could have

Gmail · Outlook · Google Calendar · Google Drive · Instagram · Facebook · LinkedIn · LinkedIn Page · X · Slack · Trello · Notion · Shopify · WooCommerce · PrestaShop · Stripe.

The selling point isn't the count. It's what we didn't ask for:

• 9 Google scopes cut to 5. We dropped mail.google.com entirely.
• Google Drive runs on per-file access only. We swapped the "read all your files" scope for drive.file plus the official Google Picker. You choose the file; the agent never sees anything you didn't personally hand it.
• Shopify is read-only by grant - all seven scopes are read_*.
• Stripe is read-only by construction. Only get_ and list_ tools exist. There is no charge, refund, or payout code path in the product to misuse.
• Deliberately refused: LinkedIn's email address scope, Slack's file read, Trello's account scope, and X's DM, follow and like write scopes.

All 16 integrations →

Automations - work that happens while you're offline

Ask any agent for something recurring, in plain English, and it builds a real scheduled job that runs as that same agent and leaves a draft waiting for you.

The "don't send them" is redundant. It can't.

What automations do: run on schedule, use the full toolbox, prepare the work, notify you. What automations never do: send, publish, or post unattended. The schema accepts an unattended setting; the code downgrades it to a one-tap approval card, and the runner refuses to send on its own.

Honest limits worth knowing before you plan around them:

• Schedules are daily at a time, or an interval of 30 minutes or more. No cron expressions, no weekday picker, no monthly rules.
• Automations are created by talking to an agent. The Automations hub lets you pause, resume and delete them - not write them.
• New accounts start with zero automations. There's no template pack. You build the ones you want, in a sentence.

How automations work →

Tasks - the shared to-do list your agents can read

Any agent can open a task in your Neural Core with a title, a description, a due date and an owner - and a task assigned to a coworker shows up in that coworker's pending work the next time you talk to it. You manage the list from the sidebar or the Neural Core.

Ask Ema "what's on my plate?" and she runs a single cross-agent sweep: unread notifications, posts scheduled or failed, emails still pending, automations coming up.

Tasks →

Marketing Studio - a place to start, not a blank page

The Studio is the create-and-browse surface: an inspiration gallery you can scroll, and a one-click "recreate this" that drops the prompt straight into the composer for Tyron to make your version.

For images, the real workhorse is the conversation. Ask Tyron for an image in any aspect ratio, then keep talking to it - "make it warmer", upscale it to HD, cut out the background - all in the same thread, all on the same picture.

Marketing Studio →

Website builder - describe it, watch it get written, publish it

Steve is the tenth agent and the most watchable one. Describe a site and the code streams into a live file tree while it's being written - real code, typed in front of you, with a working preview you can click.

• Click any element in the preview to tell Steve exactly what to change. The inspect script that makes this work is injected into the served preview only - never into your ZIP, never into your published site.
• Edits are surgical, and every one lands as a red/green diff in the transcript.
• Publish to a real public URL at your-name.unyo.site, with a certificate, in one sentence.
• The preview never dies. If the sandbox is gone, it's rebuilt for free from the durable copy of your files.
• Export a ZIP and self-host it anywhere.

What Steve builds: static front-end sites - HTML, CSS, JavaScript, a CSS framework over CDN. What Steve does not build: anything with a build step, npm, a bundler, a backend, a database, logins or API routes. Caps: 40 files, 512 KB per file, 12 MB per project.

Pricing: build 25 credits · edit 15 · publish 50 · republish 25 · add an image 5 · ZIP export 10. Reading your code, unpublishing and restarting a preview are free.

The website builder →

Voice - talk, and watch the work appear

Open a call and talk to your agent. While you speak, the work shows up as cards in the conversation behind the call. Eleven languages.

Voice is deliberately narrow: nothing publishes or sends by voice. The card appears, the call ends, you tap. A spoken "send it" is not a signature.

Voice →

Skills - 33 one-click workflows

For the jobs you do the same way every time, there's a page instead of a prompt. 33 purpose-built skill pages across four agents - Ashley for social, Tyron for media, Maya for email, Alex for sales. Trend finders, caption rewriters, best-time-to-post, hashtag research, background removal, image enhancement, cold outreach, objection handling, and more.

Riley, Sam, Lucy, Blake, Ema and Steve are chat-only today. That's not a gap in what they can do - it's a gap in shortcuts to it.

Browse the skills →

Documents, spreadsheets and real analysis

• 9 of the 10 agents write real documents - article, brief, report, press release, product sheet, newsletter - delivered as an editable card, never pasted into the chat as text.
• Real inline [n] citations that survive editing and reloading, with a references appendix localised in 10 languages.
• Every agent can build a real multi-tab .xlsx you can open, export to CSV or PDF, or push to Google Sheets. No sandbox needed.
• Blake runs real Python - pandas, numpy, matplotlib - over your data. He doesn't estimate arithmetic, he computes it, and hands back charts and downloadable files.
• Attach a PDF, DOCX, XLSX or CSV to any message and ask any agent to read it. Scanned PDFs still work - OCR takes over when there's no text layer.
• Deliverables live in a private, owner-only bucket. No download link is ever stored.

Publishing & scheduling

• Real native carousels on all four platforms, each built through that platform's own API flow.
• Publishing is idempotent - a fingerprint of the content is the job key. Click twice, get one post.
• Advice from your numbers. Best-time-to-post and hashtag rankings read a cache of your own posts' performance, not invented industry benchmarks.
• Schedule, reschedule and cancel in a sentence - "move it to Friday 10am".

Security - the part that lets you say yes

The boundaries above only mean something if the foundation holds. As of 16 July 2026, verified by live query:

• 100% row-level security coverage - 109 of 109 public tables, 291 policies. Zero unprotected tables.
• 17 of 17 populated OAuth token columns are AES-256 encrypted at rest, verified at the cipher-byte level.
• The master key is unreachable by any client role. It lives behind a function granted to one internal role and nothing else.
• Zero cross-tenant access surface - a live audit for privileged functions that trust a caller-supplied user id returns nothing.
• Independent CASA Tier 2 penetration test (TAC Security, 6 May 2026): 0 Critical, 0 High, 0 Medium.
• 17 irreversible actions are confirmation-gated by database configuration, not by prompt. A model that misspells a tool name cannot slip a send through - the fuzzy-name guard refuses to auto-run anything that costs money, needs confirmation, or has an external effect.
• Retention is a running job, not a promise: 03:00 UTC daily - conversations at 90 days, cached emails at 30, logs at 7.
• Deleting your account purges roughly 80 tables atomically, revokes every provider that offers a revocation API, takes your published websites offline first, and completes even if a provider is unreachable.
• Every model runs under a no-training guarantee.

The security page →

Pricing - you pay for work, not for talking

• 2 500 free credits at signup. No card.
• Billed on real cost, not per message. Credits are computed from what a run actually cost to run.
• 73 of 105 priced tools cost zero credits - every read, every search, every draft, every plan.
• Failed work is refunded automatically. All 32 paid tools refund on error, and an hourly janitor refunds anything left stranded.
• The credit gate fails open on infrastructure trouble and closed only on a genuinely empty balance. A slow balance check never blocks a legitimate message.
• Top-up credits never expire, and they're spent before your monthly quota.

Plans: Starter $39 / 25 000 credits · Pro $99 / 80 000 · Business $249 / 250 000. Top-ups: $10 / 4 000 · $25 / 12 000 · $50 / 30 000. EUR at parity.

Full pricing →

Languages

11 agent reply languages · 6 interface languages · 11 voice languages. Set your agents' language once and they answer in it no matter what you type in. Your interface language is a separate setting - changing what Ashley says never changes what your buttons say.

Watch the work happen

Every run is visible. The agent declares its plan up front, announces each step as it starts, and each step flips from pending to in progress to done or failed in front of you. Expand a finished step and you read what it actually produced - the real caption, the real subject line, the real finding.

And what you see streaming is what you see after a refresh. Live and reloaded transcripts are built by the same code over the same events, so the two can't drift apart.

Close the tab mid-run and the work keeps going server-side. Come back and it's waiting.

If an agent needs a tool you haven't connected, it pauses instead of failing and shows you a Connect button. Connect it, and the run resumes from exactly where it stopped - no retyping.

FAQ