Chat
The console's Chat page lets every member talk to any text model on the network, with an agent that can search and read the web. It runs on your gateway and pays like any other request: through an API key, within its budgets, and recorded in Logs.
Who pays
The first time you open Chat in a workspace, the gateway creates an API key for you there, labelled Web chat. Every chat request uses that key, so:
- It pays from the workspace's wallet.
- Your member budget, the key's budgets and the workspace budget apply. See Budgets and spend limits.
- Your routing policies apply, and the model list follows them.
- Each request shows up in Activity and Logs under the Web chat key.
Below the message box, the page shows which key pays, what it spent this month (against its monthly cap, if any) and an estimate of what the current chat cost. The estimate uses the provider's advertised prices; the amount actually paid is in Logs.
An organization admin who opens a workspace they are not a member of gets an unowned key labelled Web chat (<their name>).
Pick a model and a provider
- Choose a model in the model picker. Models with free offers are marked Free offers; others show their price range.
- Choose a seller, or keep Auto: the gateway picks the best offer each turn. With Auto, the page shows which provider automatic routing would pick right now.
- Type your message.
Every answer shows which provider served it, the tokens used and the estimated cost. Pinning a provider sends every turn to it. If it stops offering the model, the page offers Use Auto.
If no free provider offers the model and the wallet is empty, the page links to Add funds. See Funding.
Tools
On models that can use tools, the agent can:
- Search the web and read web pages.
- Read and write files in a private scratch directory for the conversation, on the gateway host.
Models that cannot use tools answer as plain chat. Web search uses Brave Search when the gateway has ANTSEED_BRAVE_API_KEY set, and DuckDuckGo otherwise.
Shell commands (organization admins)
Organization owners and admins also get a shell tool. It runs commands in the chat's scratch directory, as the gateway's own operating-system user, with a 60-second timeout.
No command runs on the model's say-so. Each one stops the answer and asks Run this command on the gateway host?, showing the exact command, with Approve and Deny. Approve only commands you understand: the model's output comes from a network provider and from any page it read, either of which can try to steer it.
A shell command has the gateway user's permissions, including access to its data directory and wallet keys. The scratch directory confines the file tools, not the shell.
Conversations
Your conversations are stored on the gateway and listed on the left of the page. Each member sees only their own. Delete one from its row.
Chat from the terminal
antseed chat opens the same chat in your terminal. It uses the same backend, agent, tools and stored conversations as the console.
On the gateway host, no key is needed: the first run creates a key labelled CLI chat in the Default workspace and keeps it in the data directory.
antseed chat
antseed chat --workspace Research --model deepseek-v4-flash
Against a remote gateway, pass its origin and an API key:
antseed chat --gateway https://gateway.example.com --key "$ANTSEED_API_KEY"
| Option | Purpose |
|---|---|
--gateway <url> | Gateway origin (default: the gateway on this machine) |
--key <apiKey> | API key that pays, also read from ANTSEED_GATEWAY_KEY |
--workspace <id|name> | On the gateway host: the workspace the created CLI chat key pays from |
--model <id> | Model id (default: the most offered model that can use tools) |
--peer <peerId|any> | Provider for every turn; any lets the gateway pick each turn |
--resume <threadId> | Continue a conversation |
--history | List your conversations and exit |
A chat run with an API key never gets the shell tool.