Experiential Agent
Chat
One place to use any model in the catalog. Ask a question and get an answer, or hand a model that can use tools a whole task and get finished files back. Compare models side by side and copy the code that makes the same call from your app.
Open Chat
Choose Chat in the sidebar. The page opens on a model picked for you. Use the model name at the top to switch to any other model, type a message, and press Enter. Shift and Enter start a new line.
Every message is a real call through the gateway, billed to your organization's credits at the model's normal price. Nothing here is a demo tier. The same model answers the same way from your code.
The model decides what Chat does
There is no switch to flip. Pick a model and Chat works the way that model can. Every model in the catalog can be used here.
| The model you pick | What Chat does |
|---|---|
| A chat model that can call tools | Works as an agent: its own computer, your files in and finished files out, runs code, searches the web, and keeps working if you close the tab. |
| A chat model that cannot call tools | Chats: you ask, it answers. |
| An image model | Makes images from your prompt. |
| An embedding, speech or transcription model | Opens that model's own mode: vectors, spoken audio, or text from audio. |
Every chat is one conversation, whatever models answer in it. The model you pick answers your next message, and you can switch models at any point without leaving the conversation: the new model reads what came before. With a model that can call tools, the agent works on your message on the conversation's own computer: it plans, uses tools, and shows its progress as it works. You can steer it, stop it, and start another conversation while the first keeps going. See Agent for everything the agent can do and its limits. To ask a quick question without tools, pick a model that cannot call them, or just ask: the agent answers directly when no tool is needed.
A good first chat
- Pick a model from the model name at the top.
- Paste a short email you are not sure how to answer and ask for three different replies.
- Attach a spreadsheet and ask for a one-page Word summary with a chart. A model that can call tools hands back the file.
- Open Settings and read the Last response panel for the tokens, cost and time of a reply.
Compare models
Choose Compare in the toolbar to add a second model. Each model gets its own column and every message you send goes to all of them at once, so you can judge the answers, the speed and the cost on your own prompt. You can compare up to four models. Use Add model for the next one, and the X on a column to remove it.
Settings
Open Settings with the panel button at the top right. It shows the selected model's context window, maximum output and prices, then only the controls that model supports. A control the model does not support is not shown, so you cannot send a setting the model would refuse. A control you never touch is not sent, and the model uses its own default.
| Setting | What it does |
|---|---|
| Temperature, Top P, Top K | How adventurous the answer is. Lower is steadier and more repeatable; higher is more varied. |
| Max tokens | The longest reply you will accept. A token is roughly three quarters of a word. |
| Reasoning effort | How hard a reasoning model thinks before it answers. Higher effort is slower and costs more. |
| JSON response | Asks the model to answer as a JSON object, for when another program will read the reply. |
| Stop sequences | Text that ends the reply as soon as the model writes it. |
| Logprobs | Shows how likely each word was, for people studying a model's behavior. |
| Tools | A list of tools you describe in JSON, so you can test how a model asks to call them. |
The Last response panel shows the input and output tokens, the cost and the time of the conversation's latest chat reply.
Files and previews
Attach files with the paperclip or drop them into the window. A model that can call tools works on copies of any file on its own computer, so spreadsheets, documents, PDFs, CSVs, images, audio and archives all work, and they stay in the conversation's files. A model that only chats reads attachments with your message instead, the kinds it can read (for example images for a model that can see), and the composer tells you which; a conversation that only ever chats never starts a computer. Attachments are available when you talk to a single model. You can also paste: a screenshot or files you copied (Cmd or Ctrl+V in the message box) attach the same way, with a thumbnail for images.
Files the agent makes appear under its answer and in the Files panel. Click one to preview it: documents, spreadsheets, slides and PDFs preview as pages, and images, charts, audio, video and text show in place. Download one file, or use Download all for a zip.
Edit and resend
To change something you asked, hover over your message and choose the pencil. Edit the text and send it (Enter sends, Escape cancels). Edit is not shown while a reply is streaming or the agent is working.
- When every reply from that message on is a chat reply and you send the edit to a model that chats, the message and everything after it are replaced in the same conversation, like any chat app. In Compare, every column is cut at that message.
- When the agent worked on that message or a later one, its computer cannot be rewound: editing your last message sends it again in the same conversation, and editing an earlier one starts a new conversation from the edited text. The editor tells you when that happens before you send.
Regenerate
Hover over the latest answer and choose Regenerate to send your last message again, or the small arrow beside it to pick another model first. A chat answer is replaced by the new one (the earlier answer is not kept). When the agent worked on that message, it gets a new turn with the same message, because its computer cannot be rewound.
History
Choose the clock button at the top left (or press Cmd or Ctrl+K) to see your conversations, newest first. A spinner marks one the agent is still working in. Open any conversation to continue it. The ... menu on each conversation pins it to the top, renames, shares or deletes it. New chat starts a fresh conversation.
The search box at the top finds conversations by their title and by what you wrote in them. Replies are not searched. Escape clears the search, and a second Escape closes History.
Conversations are saved with your organization, for you only: open one on any device and it picks up where it is. Chats an earlier version kept in your browser move into this list the first time you open Chat. Compare and the image, embedding, speech, transcription and video models are tools for trying something out: their results are not saved.
Share a conversation
Choose Share (in a conversation's menu, or the share button at the top) to make a read-only link. Anyone signed in to your organization can open it; nobody outside your organization can, and there is no public link. The link shows the conversation as it was when you shared it, with its images; your later messages appear once you choose Update link. The agent's computer is never shared. Its saved files can be downloaded from the link only if you tick Include files; otherwise people see file names only. Unshare turns the link off at once, and deleting the conversation removes it too.
Someone who opens your link can choose Continue in a new conversation to get their own copy to keep talking in. Your conversation does not change.
What each reply cost
Each reply shows what it cost, and the composer shows the conversation's total so far. Hover over a reply's cost to see how long it took to start answering (and, for the agent, to take its first action). Each kind of spend is listed by name:
| Shown as | What it means |
|---|---|
| $0.0123 | Credits charged to your organization. |
| $0.0123 BYOK | Calls on your own provider key. The amount is that provider's list price, and the provider bills you for it. |
| Struck through, $0.0123 free | Calls a promotion or your plan paid for. Nothing is charged. |
See Credits and billing.
Use in code
Use in code in the toolbar opens a panel with the call for the selected model and your current settings, in cURL, Python and JavaScript. Copy it into your project. The key is shown as the $EXPLABS_API_KEY placeholder; your real key is never put on the page. See Quickstart and Authentication.
Run it on your computer
The exp-code button at the top right shows how to run the same open-source agent in any folder on your own computer, from your terminal, with your API key. See Run it on your computer.