An office of AI agents that runs on your computer, keeps its own records, and asks before it acts.

An agent is more than a system prompt. It has a name and a one-line job, a plain-language brief you can edit, its own conversations, the tools it may use, a budget per conversation, a model tier, memory files, scheduled recipes, and optionally an email address of its own. Hire one with a sentence; the app drafts the brief, gives it the team's shared memory and the folder you are in, and you read and change the rest on its page.

Agents work together. One can hand a task to another, and the result comes back as a card: who did it, how long it took, what it cost, and Open to read the whole thing. Nothing opens a new window. A hand-off is itself something you approve, so an agent cannot quietly spend another agent's budget.

While an agent works you see each tool it uses as one line: a dot, the tool's name, what it touched, how long it took. Click the line for the full arguments, output and any files, with Open and Reveal. Branch a conversation from any message to explore another direction without touching the original. Retire an agent and its conversations stay; hire it back under the same name and its schedules and work come back with it.

You do not write cron expressions. Under Schedules, “Say what should run and when” takes a sentence like every weekday at 8, read the inbox and list what needs a reply, drafts the recipe, and schedules it. Any conversation becomes a recipe from the composer's menu, so a job you did once by hand can run every Monday from then on. Every recipe can be a slash command: type / in any composer and pick it.

A schedule shows when it runs next and its last runs; each last run opens the conversation it produced, so you read exactly what happened, including what it could not do and what it cost. A scheduled recipe runs with the approval mode it was saved with. An agent you have not marked as working on its own does not run unattended; its due jobs wait in Work for your Run.

Routine checks are cheap by design. A recipe can carry a pre-check, so an inbox job that finds nothing unread skips the run with no conversation and no model call. A run cut off because the app was closed says so, and runs again at its next time.

Work is the one place that looks across the whole team. Open it from the rail; the badge counts what needs you. Every row opens the conversation it belongs to, so an approval is answered where it was asked and a finished piece of work is read in context. An empty Work says so in one card and stays out of the way.

A seat's interview writes its first memory: your hours, your calendars, what counts as a decision for you, who you keep for yourself. From then on each agent keeps standing facts about you and your work as plain files in its own folder, beside the team's shared memory. You can open them with any editor, and the agent's page shows what it knows.

The Archivist keeps that memory honest. Every night it compacts each seat's dated notes into standing facts. Every week it reports what the office remembers, what it forgot, and where two seats disagree. It tracks promises one seat made that nobody kept, and hands an insight from one seat to the seat that needs it. It never rewrites a fact you stated, and it never reads the walled seats.

Agents can recall across their past conversations, so “what did we decide about the vendor in March” is a question, not an archaeology project. And memory is custody, not trust: a memory write that cites a file nothing could see is refused. A reply that calls something verified is checked against what actually ran. The agent's ledger counts those refusals, and you can read it.

How often it asks is one word in the composer's footer. Smart asks for anything hard to undo and lets reading pass. Manual asks for everything. Autonomous asks for nothing, which suits a scheduled recipe you have watched a few times more than a conversation. Talk only is a model with no tools at all. An approval waiting in any conversation also appears in Work, under Needs you.

Some things ask in every mode. Force-quitting an application, or launching a browser in a way that another program could drive, is gated even when a conversation is set to Autonomous. The seat packs add their own floors: no seat sends mail with your name on it without approval or a playbook you wrote, and no seat can move money. Tools are switched per agent, so the treasurer can reach the ledger and nobody else can.

Every run leaves a record: what it did, what it cost, what it could not do. Each agent has a budget per conversation; when it is reached the agent stops and says so, and Work shows the stop. Provider and billing errors surface in the conversation. Nothing fails silently.

Email. Give an agent an address and mail to it shows in Work as theirs. With Gmail connected, the Chief of Staff reads your mail for decisions and commitments while the Assistant answers it: replies land in your Drafts folder in your voice, and you send. Hand the Assistant a playbook and it can run a second desk inbox on its own, within the rules you wrote.

Texting. Turn on Texting under Settings, open the Cluer bot in Telegram, send the six-character code the app shows, and your agents answer on your phone. Each agent keeps its own conversation there; /agents lists your team and /name switches to one. The bot belongs to Cluer, so there is nothing to set up on the Telegram side.

Quick Ask. A global shortcut opens a small window over whatever you are doing. Type, press Enter, and a new conversation with the agent you choose opens in Cluer with your question already sent.

Voice. Dictation runs on the computer itself, no key and no upload. On a Mac, calendars and reminders come straight from the system, so a seat reads your day without a cloud account. Web search runs through Cluer with no key to enter. Gmail, Google Calendar and Drive connect through a guided setup where your credentials never pass through a chat.

Tools are what an agent can do beyond talking: read and write files, run a command, search the web, work a calendar, read and draft mail, drive a browser, reach your accounts. Every tool has a switch for the team and a switch per agent, so narrowing is one click. A tool that could not start says why under the conversation header, and the agent works without it and says so when it would have helped.

Skills are written know-how: the steps, what to check, what to do when it goes wrong. An agent loads a skill when the task calls for it, so the same job comes out the same way every time. The marketplace holds 190 of them; 53 general ones arrive on every install without a click, and the seats bring their own. Agents can write skills too: ask one to turn a job it just did into a skill. Sharing a skill back runs a privacy scan first; anything that looks private stops the send and says what it saw.

Stacks install a set of tools and skills for one kind of work together: a desktop or web app builder with auth, payments and crash monitoring wired in, or a mobile app builder for iOS and Android. Apps are small things an agent built for you and kept: a form, a tracker, a report that refreshes. Each opens in its own window, and Export takes one to another machine.

Left alone, Cluer picks the tier turn by turn. The pill in every conversation header shows which tier answered and the running cost. Click it to hold a conversation on one tier; an agent has a default on its page, and a seat can be pinned so it never drops below the tier its work needs. When a different tier would clearly do better, the pill grows a dot and its menu says so: Switch or Stay.

Settings → Plan and usage is the money page. One number at the top, this month's spend; under it a bar shows what is spent, what the plan includes, and the room left before the limit you set. The monthly limit is a dial, not a rule the app makes for you. Work stops only when it is reached, and you can raise it any time. The daily guard pauses a runaway day so you can look before continuing. This month by model shows where the spend went. Manage card and invoices opens your billing in the browser.

New models reach your install without an update: the tier table is maintained centrally and refreshed daily, with a human deciding each swap. When a month runs past what your plan includes, the Plan page shows whether Plus would have cost less. It says nothing when it would not.

Conversations are in a local database in the app's data folder. Your team, each agent's page and their memory are files in the instance folder beside it. Keys and account tokens are in the system keychain, never in a file. Work, ledgers, journals and the seats' records live in folders you can open with any app.

Export, under Settings → Advanced → Data, builds one backup file of everything above; Import restores it on another machine. Removing the app does not remove the data folder. Your own conversations also follow you between your machines: each device pushes what it added and pulls what the others added, so a merge never overwrites a conversation that only one machine had.

A shared folder is a folder on your disk that your agents can read and write, and that syncs encrypted between the machines and people on your team. It is where the draft, the spreadsheet and the notes live, so an agent's work lands somewhere you can open with any app. Agents create shared folders when you ask; the drawer shows the path, when it last synced, and Sync now.

A scope is a permission boundary for shared work: the folders, tools, network hosts and budget an item in it may use. Only you can make one. Invite colleagues by email; each joins with their own sign-in, and nothing from the scope runs on their computer until they approve exactly what it may reach there. If the owner widens a scope later, nothing runs until they approve again. Credentials are never in scope; there is no field for them.

An agent can ask for a scope, too. The request lands in Work with its ceiling spelled out, and you change anything before you approve it. Items in the queue carry text as data, never as authority: a task cannot widen its own permissions by asking nicely. Agents on other platforms can take work from the same queue over plain HTTPS with a token pinned to one scope.

When an agent leaves the team, its schedules pause and its items say so in Work; hand them to someone on the team and they run again. A collaborator you bring in can find an agent already waiting at first sign-in, built for them by the owner, with its brief, its skills and its starting files in place.