A chat assistant for research and longer tasks.
An assistant on the models you connected. It searches and reads the web, works through your files, hands pieces to subagents, and keeps running on the server after you close the tab.
Free includes 25 web searches and 25 page reads a day.
Compare the three most cited benchmarks for retrieval over long PDFs, with sources.
Web on: a search returns six results, and it reads the pages behind them. New chats start this way.
Watch a thread from request to answer.
A request goes in. The agent plans, searches, reads and writes, and you watch each step land in the thread as it happens.
Web research with sources you can open.
Every step lands in the thread as it runs. Open the research agent to see what it was asked and what it read, or open the sources to check the answer yourself.
Compare the three most cited benchmarks for retrieval over long PDFs, with sources.
Three benchmarks come up in almost every paper on the topic arxiv.org. They differ most in how long the documents are and in whether a question needs a table or a figure to answer huggingface.co.
If your PDFs are scanned reports, weigh the third one highest. It is the only one scored by a judge.
- SearchSix results a call, or ten with Analysis on. Indexed sites you crawled are searched before the open web.
- Read a pageThe readable text, up to about 20,000 characters. A page that comes back nearly empty is reported as blocked, not as nothing.
- Research agentTakes the whole question and runs its searches and reads side by side. About twelve steps shallow, twice that deep.
- SourcesAny answer built from the web lists every page it used, with its site, title and description.
Five modes that set how much work the agent does.
The mode sits next to the plus button. It decides how long the agent may work, how many subagents run at once and which tools it may touch. It never changes which model answers.
Normal
Everyday work
- Steps a turn
- 300
- Time a turn
- 45 min
- Subagents at once
- 2
- Plans
- Every plan
The default for everyday work. Two subagents run side by side, and attaching a Computer raises that to five.
Effort is its own setting in the same menu:LowMediumHighMax
Close the tab. The job keeps running.
Sending a message queues a job on our servers, not in your browser. It keeps running whether the tab is open or the laptop is asleep.
- It replays what you missedReopen the chat and it reattaches to the running job, then replays everything that happened while you were gone. The live replay is kept for an hour. After that the finished message is still there.
- It pauses when it needs youIf the agent needs an answer from you, it waits 90 seconds and then suspends instead of holding a worker open. Pick it up again with /resume.
- It queues when we are busyA deep queue shows your position. From Pro up, chat jobs go ahead of free traffic, and nothing you send is lost while it waits.
- Live streaming stays on one deviceThe pointer to a running job lives in the browser that started it. Another device sees the answer once it is saved, rather than watching it stream.
Artifacts open in a panel beside the thread.
When the assistant makes something you will want to keep, such as a document, a chart or a web page, it becomes an artifact with a preview, an edit view and a version history.
- Editing makes a new version instead of overwriting. Setting an older one as current forks forward from it, so nothing after it is destroyed.
- The assistant edits by replacing exact strings rather than rewriting the file, so your own changes elsewhere survive.
- A website previews at phone, tablet and desktop widths, with an inspector and a console that counts errors.
- Something too long for one pass goes to a build subagent that drafts it section by section.
Seven kinds, each rendered as itself
- Bakery landing pageView websiteA live page in a sandboxed frame
- Onboarding notesView documentFormatted text
- retry.tsView codeHighlighted and editable, never run
- Logo sketchView imageThe image, over light, dark or a checkerboard
- Signup flowView diagramA Mermaid diagram you can pan and zoom
- Signups by weekView chartA line, bar, area or pie chart
- Q3 board reportView documentThe document itself
Attach large files. Long ones are indexed and searched.
Up to 40,000 characters goes in whole, as if you had pasted it, with up to 60,000 across the files of one message. Anything longer is indexed and searched, so a 200 page PDF costs about as much context as the paragraphs that turn out to matter.
Where does the report say the error budget comes from, and does the code agree?
Images skip the index and go straight to a model that can see. Attaching a file you already sent reuses its index and does not count against your upload quota.
- 10 MBa file
- 5documents a message
- 10images a message
- 3repositories a chat
Define your own agents. The assistant remembers your preferences.
Save a worker once and call it by name, or let the assistant pick it from its description. What you tell it about yourself carries into every conversation after.
- 1GenerateDraft the notes from the merged pull requests.
- 2ValidateEvery entry links its pull request. Returns pass or fail with a reason.On fail: retry, up to 5 times
- Up to twelve steps, each with its own prompt, model and tools.
A custom agent has its own system prompt, tools, connectors and, if you want, its own model and effort. Run it as one pass or as a pipeline where a validate step judges what a generate step made. Custom agents
Memories
- Wants Python examples typed, with dataclasses rather than dicts
- Writes in British English
- Deploys the API to a single server in Frankfurt
- Prefers short answers with the command first
A memory is saved by the assistant when you state a lasting preference. Read, edit or forget any of them. What you say in the current chat always wins over a memory. Memories
Connect databases, servers and apps. Writes need your approval.
Reads run without a prompt. Anything that could change something stops and asks first, and commands marked dangerous can only be approved one at a time.
Credentials are sealed and only decrypted for the length of one call. An allow list entry belongs to one chat and one connector, and never carries over to another conversation.
This query writes to orders-db. Run it?
update orders set status = 'shipped' where id = 4120;
In the chat, credits pay for search and page reads only.
Your provider bills every token the model reads or writes. Outside the Computer, a chat draws credits for three things, and only on a paid plan.
- Web search€0.006a search
- Page read€0.003a page
- Deep search€0.05a run
Results are cached for a day, so asking the same thing twice does not cost twice. When a daily limit is reached, the assistant stops searching and reports what it already found.
| Limit | Free | Starter | Pro | Max | Enterprise |
|---|---|---|---|---|---|
| Credits each period | none | €10 | €25 | €90 | unlimited |
| Web searches a day | 25 | 1,000 | 5,000 | 15,000 | Fair use |
| Deep searches a day | 3 | 200 | 1,000 | 3,000 | Fair use |
| Page reads a day | 25 | 1,000 | 5,000 | 15,000 | Fair use |
| Custom agents | 3 | 10 | 30 | 100 | 100 |
| Uploads a day | 5 | 100 | 300 | 600 | Fair use |
| Uploads a week | 20 | 500 | 1,500 | 3,000 | Fair use |
| Images a day | 25 | 100 | 300 | 600 | unlimited |
| Images a week | 100 | 500 | 1,500 | 3,000 | unlimited |
| Image storage | 1 GB | 10 GB | 50 GB | 200 GB | unlimited |
- Credits each period
- €25
- Web searches a day
- 5,000
- Deep searches a day
- 1,000
- Page reads a day
- 5,000
- Custom agents
- 30
- Uploads a day
- 300
- Uploads a week
- 1,500
- Images a day
- 300
- Images a week
- 1,500
- Image storage
- 50 GB
Free has no credit balance, so nothing on it is charged; it uses the daily limits above instead. Images run on your own image model, billed by its provider. Every limit on every plan
Get started
The gateway, the chat and the Computer share one account, one set of API keys and one credit balance.