Agent

What the agent can do

• Read the files you upload, including spreadsheets, documents, PDFs, CSVs and images.
• Keep a visible task list so you can see the plan and its progress.
• Write and run code to calculate, clean data, convert formats and draw charts.
• Search the web and read pages, and cite where its facts came from when you ask.
• Ask you a question when a request is genuinely ambiguous, instead of guessing.
• Hand back files: Word documents, Excel workbooks with live formulas, PowerPoint decks, PDFs, charts, CSV and Markdown.

Start a task

1. Open Chat in the sidebar and choose a model that can call tools.
2. Describe what you want done. Use the paperclip, or drop files into the window, if the agent needs them.
3. Press Enter. The agent starts work and shows its progress as it goes.

Good first tasks

The empty screen offers four starters you can click. They are a good place to begin:

• Report from a spreadsheet. Attach a spreadsheet and ask for the key totals, the trends worth knowing, and a one-page Word report with a chart.
• Make a slide deck. Ask for a six-slide PowerPoint introducing your product to a new customer.
• Research brief with sources. Ask for a short brief on a topic with a source link for every claim.
• Clean up a messy CSV. Attach a CSV and ask the agent to fix dates, remove duplicates and tell you what it changed.

A good request says the outcome you want, the audience, and the format. "A one-page summary for my manager as a PDF" works better than "summarize this".

Watching it work

While the agent works, its current action stays in view and earlier steps fold away. Its task list is pinned to the reply and ticks off as steps finish. When it is done, the message shows your request, the task list as one line (for example "5 of 5 done"), the work folded into a single line you can open to see every step, the answer, and the files it made.

Steer, stop, or start another task

• Steer. Type a message while the agent works. It reads it after its current step and changes course, for example "use last quarter, not this quarter".
• Stop. Press Stop to halt the current step and any command it is running. Your next message is treated as a new request, not a continuation of the stopped one. If the agent was interrupted or ran out of steps, Continue picks it up.
• Start another conversation. Choose New chat in the toolbar at any time. The agent keeps working in the first one in the background.
• Answer a question. When the agent asks, pick one of its options or type your own answer.

The agent keeps working

The agent does not need your tab. Close it, switch devices, or reload, and the work continues. Open History (the clock button at the top left) to see every conversation, with a spinner on the ones the agent is still working in. Open any conversation on any device and it picks up where it is. You can rename a conversation or delete it from the same list.

A conversation can mix models: switch to a model that only chats for a quick question and back to the agent in the same conversation. The agent reads the chat replies it did not take part in before it starts on your next message.

Files in and out

Upload files with the paperclip or by dropping them into the window. The agent works on copies on its own computer; your originals are untouched. Choose Files to see everything the conversation has produced and uploaded. Click a file to preview it (documents, spreadsheets, slides, PDFs, images, charts and text all preview), download one file, or use Download all to get every file as one zip. Large tables preview their first rows and large text previews its first megabyte; download the file for the rest.

What it costs

Every model call and every web search the agent makes is billed to your organization's credits at the model's normal price. The computer it works on is not billed. Each message shows what it cost as it accrues, and the composer shows what the whole conversation has cost so far: $x for credits charged, $x BYOK for calls on your own provider key (billed by that provider), and a struck-through $x free for calls a promotion or your plan paid for.

Each run has a spending limit, so a task that goes wrong cannot run up an unbounded bill. If your balance runs out, the task stops with a link to add credits. See Credits and billing.

Which models can drive the agent

Any chat model whose lead provider supports tool calling. Other chat models chat without tools, and image, embedding, speech and transcription models keep their own modes. A model that reads images lets the agent look at images you upload; for other models the agent can still process those files with code.

Limits

A conversation whose computer was cleaned up is rebuilt from its saved files on your next message, so you lose nothing you saved. Files over 50 MB are kept only while the computer is alive.

Privacy and your data

A conversation stores your uploaded files, the files the agent produced, the messages, and a log of what the agent did. Deleting a conversation from History deletes all of them, and the computer it ran on is shut down. Deleting your account or organization removes your conversations the same way. Only you can open your conversations.

Run it on your computer

The agent is open source: exp-code, published on npm as @experiential-labs/exp-code. The exp-code button at the top right of Chat shows these steps. It runs in any folder from your terminal, with the same tools, document skills and models, and needs Node.js 22.19 or newer.

1. Install exp-code.
2. Run exp-code login. It opens your browser, you pick an organization and approve, and the key is saved on your computer.
3. Start exp-code in the folder you want it to work on.

npm install -g @experiential-labs/exp-code
exp-code login
cd my-project
exp-code                                             # interactive terminal UI
exp-code -p "Summarize sales.csv into report.docx"   # one task, print the answer, exit

On a machine without a browser, exp-code login --no-browser prints the link to open elsewhere and lets you paste a key. You can also skip the login and set EXP_API_KEY to a key from the API keys page; the environment variable wins over a saved login. exp-code login --status shows which key is in use and exp-code logout forgets the saved one.

To try it without installing, run npx @experiential-labs/exp-code. If you use the Python CLI, pip install experiential, then exp login once and exp code in the folder: it signs in with your login instead of an environment variable.

Billing is the same as in Chat: every model call goes through the gateway and is billed to your organization's credits, or to your own provider key for models you connected, and each turn shows its cost. It can also run against a gateway you host yourself: start exp in another terminal and set EXP_GATEWAY_KEY to the key it prints, and exp-code calls http://127.0.0.1:8000/v1 instead (with EXP_API_KEY also set, add EXP_CODE_GATEWAY=local).

Related

• Chat: the one surface where you chat with any model, compare models, and give tasks to the agent.
• Data controls: what the gateway stores for your API calls.