Agents
What a TokenFlux agent is, where it runs, which models it uses and how it is billed
A TokenFlux agent does work for you. It runs code in a private sandbox, uses tools and the web, keeps files in its own workspace, and can run on a schedule. You talk to it in the web app or from your chat apps.
Your agents
Open Agents in the web app. On your first visit, TokenFlux creates a default agent for you. Creating it does not start a sandbox and does not spend money.
You can create more agents with New agent in the agent switcher at the top of the sidebar. Give each agent a name, a default model and, if you want, a first task. The agent starts the first task as soon as it is created.
Under the switcher, Manage opens the pages of the selected agent:
- Automations: prompts that run on a schedule. See Automations.
- Channels: Telegram, WeChat, Feishu and Slack. See Channels.
- Settings: name, default model, system prompt, and Delete agent.
Conversations lists the sessions of the agent. See Sessions.
Sandbox and workspace
Each session has its own container for the agent's file and shell tools. All sessions of one agent share one workspace, so a file that one session writes is available to the others. Sessions do not share their conversation.
An agent can also remember facts, start sub-sessions and manage its own automations. See Tools and memory.
A session sleeps after 10 minutes without activity. Its container is removed, but the workspace files and the conversation stay. The next message starts a new container.
Sessions of one agent share files, so they are not a security boundary. Use a separate agent for work that must not see another agent's files.
Models
Agents can use the models that your agent key can list, except image generation models. Choose the default model in the agent's Settings. Each session can pick its own model and thinking level. Image input and thinking levels are available only for models that support them.
Billing and usage
Agents and your own apps use one account and one balance. Agent model calls go
through the same gateway at https://tokenflux.ai/v1 as your own API calls.
All your agents share one dedicated API key, named Agent- followed by an ID.
TokenFlux creates this key on your first visit to Agents. The key has no
separate quota or expiry and spends your account balance. In
Usage, each request shows its key, so you can
tell agent usage from your own apps.
You need a funded balance and an available model before an agent can answer. Agent usage is billed per model call, at the prices on Models & pricing.
If you disable or delete the agent key on API keys, your agents stop working. TokenFlux does not replace it automatically. Enable the key again, or contact support to resolve a deleted key.
Settings
An agent's Settings page changes values for every session of that agent, starting with the next message:
| Setting | Rule |
|---|---|
| Name | Up to 64 characters. |
| Default model | Sessions that picked their own model keep it. |
| System prompt | Up to 16,000 characters. It is added to the agent's instructions for every session. |
Delete agent removes the agent, its sandbox, its workspace files, its sessions and its channel connections. You cannot undo it. If you delete the default agent, the next visit to Agents creates a new default agent with the same agent key.