The full capability set, ordered by how much it matters rather than by when it arrived.
Sorted by importance, not date. Tier 01 is what makes this a platform rather than an assistant — remove any one of those and it stops being the product. Tier 02 is what you touch on an ordinary working day. Tier 03 is the control, governance and scale that make it safe to leave running.
45 capabilities, live today.
01
The mechanisms the whole product rests on. These are the reason work gets done rather than merely assisted.
Brief a recurring piece of work the way you would brief a colleague. A planner turns it into a validated milestone plan — with concrete success criteria, scope boundaries and a cost estimate — that you read, edit and approve before anything runs.
Not everyone arrives knowing what to automate. Describe a process you run today in your own words and an interview digs out what makes it executable — the decision rules, the exception paths, the people involved. What comes back is a proposed job with every step marked automated, gated on your approval, or left manual, grounded in the systems, people and knowledge sources you have actually connected.
A job runs server-side for hours, days or weeks. It survives restarts, deploys and your holiday, and resumes exactly where it stopped. Waiting three days for a reply is an ordinary state, not a failure.
Consequential and irreversible steps are declared as gates in the plan and enforced by the execution engine — not left to a model remembering to ask. Nothing you did not authorise reaches the outside world.
A gate reaches you on WhatsApp, in the app, by email or on a phone call, carrying enough context to decide safely in seconds. You answer in your own words and the platform reads what you meant.
Every run ends in a structured review that grades what the job learned, retires what no longer holds, and applies the safe changes as a new versioned plan revision with full provenance. The tenth run is measurably better than the first, with no one editing it.
Whether a run actually succeeded is adjudicated against non-fabricable receipts of its external side effects — message ids, record ids, delivery confirmations — and against the file content it claims to have produced. An agent cannot talk its way to a pass.
Every run leaves a complete account: what was done, in what order, on whose authority, against what evidence, and what it cost. Reasoning traces, tool calls and human responses are all part of the record.
A roster of named specialists — research, planning, data, email, external correspondence, live voice, web reading, documents — orchestrated by a chief of staff. It ships finished, rather than as parts for you to assemble.
Frontier models from different vendors take the judgment calls; open-weight models take the mechanical, high-volume steps. Each seat is routed to whichever model does that job best for the least money, escalating to a stronger one when a step fails — and re-tuned as the market moves, with nothing for you to choose, learn or migrate.
Each model call is metered and attributed to the job, the run and the person who drove it, so a job's price sits next to its result instead of arriving as a monthly surprise.
02
What you actually touch on a working day — the channels you reach it through, the systems it connects to, and the things it can produce.
Google Workspace, Microsoft 365, Dropbox, OneDrive, any mailbox over standard IMAP and SMTP, and Bokio for accounting. Several accounts of the same kind can be connected under distinct names, so agents target the right one explicitly.
Describe an internal system you built yourself and agents can call it under governed, allowlisted access — credentials held server-side and never visible to the agent.
Start a job on a schedule or from an event outside the platform. Set up in plain language, with a readable URL, a readiness checklist and a way to test against a sample payload before going live.
The same conversation reachable from a browser, a WhatsApp thread or a phone call, with session memory that carries across them.
Linda answers and places real phone calls, greets each caller in the right language from the first word, and runs saved shortcuts you trigger by saying their name — asking for any detail the shortcut still needs.
A quick instruction arriving by WhatsApp, phone or email is captured as a lightweight to-do and carried out automatically — no job to set up, no ceremony.
Files, a database, a knowledge base and a memory that survive between runs, so a recurring job behaves like a dedicated worker accumulating context rather than a stranger starting fresh each week.
Genuine .docx, .pptx, .xlsx and .pdf output — including data-driven documents with embedded charts — that a recipient can open and use with no conversion step.
When a step needs arithmetic rather than estimation, the agent writes and runs real Python in a locked-down sandbox with the data libraries already installed. The script can read your project files, database collections and knowledge corpora directly — so one step can own “read all four hundred and reconcile” — and its reach is fixed by a credential minted for that single run. Binary files are transformed without their contents ever entering the model's context.
Structured data produced by one run is available to the next, editable through familiar spreadsheet vocabulary by people who do not think in schemas.
Upload your own material so agents answer from your documents rather than from general knowledge — managed without a developer in the loop.
Agree a time with several people outside your company: a professional slot-options email with one-click accept, a response matrix showing who agreed to what, and an operator calendar you can step into. Participants never need an account, and each is written to in their own language.
A plan can model the steps only a person can do as first-class milestones — a work order with an owner, a deadline and completion evidence — so one job coordinates agents and people with the same rigour.
An agent that holds a real back-and-forth with a supplier, customer or contractor over email — chasing, following up and reporting back — with your project's tone of voice and approval rules applied.
Plan a week of slots for teams, rooms or resources: pin the fixed bookings, auto-plan around them, then drag to adjust with push, swap and undo. Share by link or export to xlsx and pdf.
When something only exists on a public web page — a price, a listing, a timetable — a strictly read-only browser worker fetches the live page, without opening a door into your own network.
The research agent draws on live Google results and carries the citations through to the answer, in quick questions and deep research runs alike, so every claim can be checked against its source.
Describe the view you want over a job's real data and get a chartable widget back. It runs read-only and network-less, so it can never change or leak what it displays.
One page per job showing failure rate, cost per run, human time required and the learnings it has accumulated — plus the manual effort each run replaced, compounding over time.
Write your house rules once per project — tone of voice, audience, do's and don'ts, approval rules, glossary — and every agent, in every chat and every job, follows them without being reminded.
Durable facts about how you work are remembered and recalled automatically in later sessions, so you stop repeating yourself and never maintain a memory file.
Outcomes and progress arrive in a notification feed, acknowledged per person. The approval queue stays exactly one thing: decisions waiting on you.
03
What makes it safe to leave running, defensible to an auditor, and sellable to someone else's customers.
A monthly ceiling per account, enforced across every channel — chat, job execution, voice, triggers and WhatsApp — so nothing can run up an unbounded bill.
People hold explicit roles in each account they can reach, rather than platform-wide access, so least privilege applies and you can audit who can do what where.
A structured, searchable trail of user and agent actions, so incidents, billing anomalies and unexpected changes can be investigated without trawling server logs.
Every learning a run used is on the page — editable and deletable, with the plan hints it motivated retiring alongside it — so a job's evolution is something you steer, not a black box you trust.
Connection credentials are held server-side and never visible to the agents, which see only allowlisted operations. Data is isolated per account and processed within the EU.
Deliver the whole platform under your own brand: a hierarchy of providers with custom domains, inherited branding, host-scoped sign-in and your own agent naming — no per-customer engineering, and the reseller relationship stays invisible to the tenants below.
Invite members, reset passwords, see who last signed in, and open the product as one of your users — clearly labelled and time-limited — to reproduce a problem without asking for their password.
A secure per-account MCPA standard plug (Model Context Protocol) that lets AI tools connect to other software — the same connector everywhere, instead of one custom integration per system. endpoint exposes what a human operator can reach — jobs, chat, approvals, executions, data — so you can run and diagnose the platform from Claude Code or another external runtime without sharing a raw APIA doorway a system opens so other software can read and update its data — how the agents talk to your systems directly. key.
Use the identity your organisation already has, rather than managing another password.
Both languages are first-class across every surface — the app, jobs, live voice, transactional email, approvals and outbound messages — each reader served in their own.
Approvals, job status, results and external conversations all work properly on a narrow screen, because the person a decision is waiting on is rarely at a desk.
Work distributes across the fleet rather than cascading into timeouts: parallel milestones, chat streams that resume across servers, incremental event polling, bounded context per call and spend limits counted fleet-wide.
Join the waitlist and we'll show you what these features look like on your own work.