Chat with the built-in agent
Talk to an agent Personal Agent Proxy runs itself, with the same access settings and activity log as any agent you connect.
Chat in the sidebar is where you talk to agents Personal Agent Proxy runs itself. Every organization starts with one, General, which every member sees. An agent you make is yours alone: make one with New agent, giving it a name and a head, or ask for one from the command bar, and nobody else in the organization sees it, its owner included; you rename it, change its head or delete it. General is renamed or deleted by the organization's owner. All of them act through one entry on the External agents page, Chat, shared with everyone, as Claude or ChatGPT are one entry each however many assistants run inside them: what the agents may do with each connection is set in its column, once for all of them, and every call any of them makes is in the activity log, by the task it was made for, which says which agent it was.
Your chat with an agent is yours alone, as a direct message is, General's too: nobody else in the organization sees it. In it, the agent acts as you, so it reaches what you use: the connections shared with everyone, the ones shared with you by name, the ones you connected yourself and your own vault. Chat has a column on every connection's Permissions, beside your other agents, which whoever changes the connection's defaults changes: you on your own connections, the organization's owner too on the ones shared with everyone. On a connection shared with you by name, its owner's settings for Chat are yours too. What it does in your chat is in your activity log alone.
What it does differently from an agent you connect is where it runs: in Personal Agent Proxy, not on another company's servers. Unless you choose otherwise, the model is Claude Sonnet 5.5, on each member's own Claude account.
What runs an agent
Under the message box, a chat says what runs its agent: the model, and beside it the harness, the loop that runs its turns. For whoever may change the agent, you for one of your own and the organization's owner for General, it opens the harnesses down the side, each by its mark, with the models of the one you look at beside them to switch between. The choice is the agent's: every thread opened from then on, in any member's chat with the agent, starts with it. Threads already open keep what they run on.
A thread's own message box says what runs it, and you can switch it there, in any thread of yours, General's included: it holds for that thread alone from its next turn on, and the thread goes on from where it was. The agent, and its other threads, stay as they are.
- Claude runs on each member's own Claude account, with Claude Sonnet 5.5, Claude Opus 5.5 or Claude Fable 5.2, and counts against that member's plan. A member who hasn't connected a Claude account is told to, in the thread, and nothing of theirs runs on the server's key in its place.
- OpenRouter runs on the server's key, for every member alike, with Claude Sonnet 5.5 or GLM 5.3, each only on hosts that keep nothing of what is sent. It is only offered to an organization that has it turned on.
A new agent runs on the first harness offered, Claude where the server has it, with that harness's first model. A harness that isn't offered can't be chosen, and an agent that was set to one runs on the first that is.
Whichever harness runs an agent, each thread's title, and where it stands, its icon
and the sentence under its title, are written by a small model on the server's OpenRouter key,
gpt-oss-120b, as the command bar's reading of what you type is: it reads what the thread shows,
what you said, with the names of the files you attached, what the agent answered and the alerts
that came into it, never the calls the agent made, and only on hosts that keep nothing of what is
sent. A server without an OPENROUTER_KEY writes none: a thread is named by a line of your message,
and is Done once the agent stops, with the start of its latest answer in place of the sentence.
Threads
Say what you want done. Every message you send in the chat opens a thread of its own, about that one piece of work, and the agent works on it there. The chat lists its threads in the order you opened them, the newest at the bottom, so none moves as the agent works: each a row with an icon saying where it stands, its title, when something was last said in it, and a sentence saying where it is. The title, a few words naming the work, such as "Overdue invoices from Acme", is written once, as the agent starts on the thread; until it comes, a moment later, the thread is named by a line of your message. A thread an alert opened has a bell by its title, and keeps the name the alert gives it. A thread is:
- In progress while the agent is at work on it, a turning ring, or asleep until it goes on, a dashed one.
- Waiting for you once the agent has stopped and can't go on until you answer or do something, as when it asks whether its plan should cover one more thing: an orange ring, its sentence in full ink.
- Done once the agent has stopped and needs nothing more of you: a check.
The sentence is about the thread as a whole, such as "Merged #412 once its checks passed.", and is written as the agent's turn ends, with whether the thread waits for you, by a small model. On longer work it is written while the thread is in progress too, each time the agent tells you something and goes on, saying what it is doing and what it has found so far. Once something more is said in the thread, by you, the agent or an alert, the start of the latest thing the agent said takes its place until the next is written. A thread waiting for you has a Done button, at the end of its row as you point at it and in the title bar when it is open: press it when there is nothing more to answer, and the thread is done. A reply, or an alert that comes into the thread, takes it up again, and it no longer waits.
Several threads can be in progress at once, each on its own; everything keeps running after you leave the page, and the chat shows each thread as it is when you come back.
Choose a thread to open it beside the chat, or across the whole page, and reply in it with the box at its foot: a reply goes on with the same work, and the agent reads it with everything said in the thread before. The chat's own box always opens a new thread. A thread shows what you said and what the agent answered, and what it told you as it went on longer work; the calls it made to get there are in the activity log, not on the page.
In a thread, the agent works with your connections as the permissions say, through the same gate as any agent: a call it is not allowed is refused and logged as blocked. Each thread's work is one task in the activity log, under the thread's title and with the agent's head beside it, with every call it made under it: finished, with what the agent answered as the outcome, each time it answers, and under way again when you reply.
A thread can also wait: asked to check on something later, or to do something at a time, the agent sleeps until then, and the thread stays in progress meanwhile. It wakes with the thread and nothing else, so it says what it will do before it sleeps; a reply wakes it earlier.
The agent can also come back to you later. Ask to be told at a time, and it sets an alert, which wakes the thread it was set in then, with what it noted, so it can tell you or do what you asked. Ask to be told when something happens in your accounts, such as an email from a certain sender arriving, an issue changing or a meeting being moved, and it sets a watch on the connections it may read: each new email, issue or calendar event there that contains any of the words the agent noted comes into that thread, with a link, for it to pass on. A watch matches words, not meaning, so say what to look for the way it will be written. Alerts in the sidebar lists what is set, with the agent that set each; delete one there, or ask the agent to cancel it, in any thread: an agent is told of all your alerts, whichever thread each was set in.
Asking about your other threads
Since every message opens a thread of its own, what you ask about earlier work, such as "what is waiting for me?" or "what did we decide about the pricing review?", is a thread too. Its agent reads your other threads, with every agent of yours, through your Chat connection: each by its title, where it stands and the sentence that says so, and what was said in the ones that matter. It answers with what it found and a link to the thread, which opens beside the chat; to go on with that work, reply there. An agent never writes into another thread.
A thread can go on for as long as you like. Once it has grown long, the agent reads a summary of the earlier part in place of it, and keeps the last turns as they were; the thread shows everything said in it as before.
What you said in a chat before it was made of threads is kept as threads too: the conversation itself as one thread, Earlier conversation, and each piece of work the agent did in it as a thread of its own, under its title, opened by what you said that led to it. They read as any thread does, are Done, and a reply in one goes on from what it shows.
Unread and notifications
Each thread is read on its own: what the agent says in a thread is unread until you open that thread. In the chat, each thread whose turn ended with an answer you haven't opened has its sentence in bold and a dot at the end of its row, until you open it. The agent's row in the list counts its marked threads, and Chat in the sidebar and the window's title count them across every agent. An agent whose only news is from a thread still at work has a plain mark on its row instead.
To be told as an agent finishes, turn on notifications on the Account page: a notification of the system's, on that device, once for each thread as the agent's turn in it ends, with the agent's name and the last thing it said, not at each thing it says along the way. It opens the chat when clicked. A thread that ends while you are looking at its chat isn't told of: you see it end there. In the browser this asks the browser's leave, which it asks you for once; in the desktop app the count is on the app's icon too.
Your Claude subscription
An agent set to the Claude harness runs on your own Claude subscription rather than on the server's key, so what you say to it goes to Anthropic under your account, as it would from Claude Code, and counts against your plan rather than the server's. Connect it under Claude on the Account page; it is yours, in every organization you are in, and nobody else's turns run under it: another member's chat with the same agent runs under theirs.
The credential is a token good for a year, the same one claude setup-token makes. The desktop
app mints it for you: Connect Claude on the card opens claude.com in your browser, where you
sign in and allow Personal Agent Proxy as you would allow Claude Code; Claude shows you a code,
which you paste back into the card, and the token is made on your own computer from it and handed
to the server, which never sees your password. In the browser, run claude setup-token yourself
and paste what it prints on the Account page, or paste an Anthropic API key, which is billed to
that key.
What is stored is the credential alone, encrypted in the database as every connection's credential is, and when it ends where that is known. The server never refreshes anything, and holds no refresh token: when the year is up, the card says so and you connect again. The card also shows how much of your plan's five-hour and seven-day windows is used, as Anthropic reported it during the agent's last turn on your account, so it is as of that turn, not live.
The threads of your chat with such an agent run on the Claude Agent SDK under your account: what you say goes to Anthropic directly, as it would from Claude Code, not through OpenRouter, on the model chosen for the agent. Only where each thread stands is written on the server's OpenRouter key, from what the thread shows, as for any agent (What runs an agent). Everything else is the same: the agent's tools, sleeping, alerts, the activity log and what a thread shows. Connecting an account moves nothing by itself: an agent on OpenRouter stays there until it is set to Claude. When the token has run out or Anthropic refused it, the agent says so in the thread and the card shows it, and nothing more is asked of Anthropic with it; connecting again takes it up. When your plan's limit is reached, the agent says so, with when the limit resets, and goes on when you write again after that. Disconnect on the card deletes the credential, and deleting your account deletes it too; an agent on Claude then asks you to connect one again. The token itself can be revoked at claude.ai as well, under your account's settings.
OpenRouter
Agents on the server's OpenRouter key are being tried out with some organizations first, so the
OpenRouter harness is only offered to an organization that has it turned on, on a server with an
OPENROUTER_KEY. On a self-hosted Personal Agent Proxy, turn it on for every organization in its
database:
UPDATE "Organization" SET experiments = array_append(experiments, 'openrouter')
WHERE NOT 'openrouter' = ANY(experiments);