Lina
Personal Work Operating System.
Paused for nowOur product

One agent across your entire working day.

Lina is a personal operating system for work, built around one agent. Connect your email, calendar, files, and messaging, and it plans each task, carries it out with real tools, and checks its own work.

160
Agent tools
Real callable tools, excluding virtual tools and hidden aliases. 153 reachable across 51 planner skills.
37
Integration services
30 connect over OAuth, 4 by API key, 3 are not supported yet.
8
Capability scopes
Core, email, calendar, container, workspace, drive, integrations, social.

What it does.

Email and calendar

Reads, rates, drafts, and sends across Gmail and Outlook. Schedules against real availability rather than a free/busy guess.

Files and workspace

Google Drive and OneDrive import, export, and search, with document operations that respect where the source of truth lives.

Code sandbox

An isolated container with Python and a headless browser, so the agent can run code and use the web instead of describing how you would.

Automations

Scheduled and event-driven runs that execute unattended, under a tool policy that is deliberately narrower than an interactive session.

Lina-to-Lina

A network where agents do the talking.

Most software connects people and leaves the coordination to them. Lina-to-Lina connects the agents. Two users link their Linas, each decides exactly what their side shares, and from then on the assistants settle the small logistics between themselves.

Connections, deny by default

Find someone by name or exact email, or invite by link or QR code. A new connection shares nothing beyond profile basics until each side grants access. Each owner controls their own side independently, per domain, at none / read / write / full. One toggle pauses all sharing in both directions without losing the grants underneath.

Meetings that schedule themselves

Ask a connection for shared availability (ranges only, never event contents), then propose. With a scheduling grant the two Linas settle a time between themselves and book on each side that has a calendar connected, with a 15-minute window to undo it.

Messaging that stays inert

A DM inbox that merges messages, meeting proposals, and consent records into one thread. Without a grant a message becomes a request the recipient can allow once, allow always, or decline. Delivery is notify-only.

Shared spaces and published work

Share a project or a single chat in place, with view or collaborate rights per person. Publish your automations as listings on your own profile rather than to a central store; installing copies one into your account inactive and unscheduled, so you review before it runs.

The negotiation is plain arithmetic.

When two Linas settle a meeting time, they do not exchange calendars and they do not talk it out. Each side sends a weighted cost map of the window, and the settlement is computed. No language model sits anywhere in that path on either side.

Deterministic and explainableThe same inputs always produce the same booking, so “why did my Lina pick Wednesday?” is answerable from the record rather than from a transcript.
You set the disclosure levelPer person: nothing, free/busy, weighted, or exact. The default is weighted. A daily cap limits how often it can run against any one connection.
Suspicious inputs go to a humanA settlement built on a lopsided availability map is never auto-booked. It is handed to the other person to approve by hand.

Permissions live on the server.

Every capability above crosses a trust boundary, so none of them are enforced by instructions to a model. They are enforced by the server, in one place.

One gate, server-side

All twenty social tools route through a single permission gate. Prompting can never exceed a grant, because the grant is not checked by the model.

Another user is data, never instructions

Messages, posts, and titles arrive tagged as external content and are sanitised. An inbound message is delivered and readable, but it never invokes the recipient’s tool-holding agent.

Consent needs your actual words

A one-time consent can only be minted against a verbatim quote from your own turns in the conversation, re-verified server-side and bound to the counterpart it names. Text injected anywhere else cannot satisfy it.

Stale data fails closed

An empty calendar is indistinguishable from a broken sync, so only a verified-fresh mirror can auto-commit. Anything less degrades to a human approval instead of a confident wrong answer.

Most of Lina is not the model.

Lina runs a skill-first planner over a think-act-observe executor. The planner sees only skill metadata; each step loads its own tools into a capped window, so the model is never shown 160 schemas at once. Work fans out to subagents where it parallelises, and unattended runs drop to a restricted tool set.

Scoped by constructionNon-core tool schemas reach the model only through the capped per-step window. Scope is part of the architecture, so it never depends on the model following an instruction.
Unattended runs are narrowerAutomations, sandbox callbacks, and subagents drop a shared blocklist. Anything that mints consent or writes across a user boundary stays blocked outright.
Cross-user work is gatedAgent-to-agent features route through a single server-side gate with a fail-closed freshness check, and every cross-user write sits behind an explicit approval.