Configuring an Agent
Agents are created and tuned in Settings → Agents. You need an API key from one of the supported LLM providers; the agent uses your key to call that provider.
Create an agent
Open Settings → Agents and add a new agent. At creation you provide:
| Field | Notes |
|---|---|
| Name | The agent's name; it becomes the @handle used to mention it |
| Provider | The LLM provider: DeepSeek, OpenAI, Anthropic, Kimi, or Gemini |
| API key | Your key for that provider. It is write-only and encrypted at rest — it's never shown back, only replaced |
You bring the key from the provider's own console (for example OpenAI, DeepSeek, Anthropic, Google AI Studio for Gemini, or Moonshot for Kimi). A sensible default model for the provider is selected automatically, and the agent is created enabled.
Harness mode
With Anthropic or OpenAI as the provider, the dialog offers Use harness mode. Ticked, the agent becomes a harness agent: its turns run through Claude Code (Anthropic) or Codex (OpenAI) in a sandbox with a shell and files, on a runner that serves your workspace. The choice is fixed once the agent is created — create a second agent to have both kinds.
If no runner serves the workspace yet, the dialog says so and Create Agent stays disabled; contact support@xpt.ai to get one, or host your own. A harness agent's settings show a Harness mode block saying where its turns run; everything else is edited as below.
Tune the agent
Open the agent to edit its details:
| Setting | What it controls |
|---|---|
| Name | The agent's handle (kept lowercase) |
| Avatar | Icon and color shown next to the agent's messages |
| API key | Replace the stored key (it stays write-only) |
| Model | Which model of the provider to use. If a previously chosen model is no longer offered, the UI flags it so you can pick a current one |
| System prompt | Optional standing instructions that shape how the agent behaves — its role, tone, and rules |
| Available as workspace assistant | When on, the agent can be chosen in the AI Assistant panel (see Using Agents) |
| Max iterations | A safety cap on how many provider calls one run may make before it stops — raise it for genuinely multi-step work, lower it to contain cost |
| Enable / Disable | Turn the agent on or off without deleting it |
Changes are versioned — each settings edit bumps a revision, and each run records which revision it used, so you can tell which configuration produced a given result.
Choosing a provider and model
Pick the provider you have a key for; each provider exposes its own list of models. Heavier models reason better but cost and latency more — start with the provider's default and adjust. A "Pricing" link to the provider's pricing page is shown next to the model where available.
Multiple agents
You can create several agents with different roles — for example a planning agent with one system prompt and a quick-answer agent with another. Each has its own handle, so people choose the right one when they mention it.
A word on keys
There are two different kinds of key in the product, don't mix them up:
- Agent API key (here) — your provider key (OpenAI, DeepSeek, …) that the agent uses to think.
- Workspace API key (Settings → Connect to Claude) — a key that lets an external local agent connect into this workspace over MCP. That's covered in Work from Your Own AI (MCP).