Agents: ipx agent#
An agent is a backing model plus a system prompt and optional MCP servers, saved under your agent username and
called anywhere as agent/<username>/<agent-id>. Calls are billed at the standard per-token rates, either to the
caller or (--payer owner) to the agent’s owner.
$ ipx agent profile set acme-bot # one-time: your public agent username
$ ipx agent create support --name "Support" -m lightning --system-file prompt.md
$ ipx agent chat acme-bot/support
$ ipx agent ask acme-bot/support "How do I reset my password?"
$ ipx agent list
References are username/agent-id; agent/username/agent-id also works, and a bare agent-id means one of
yours. Agents belong to people, so management commands ignore the organization context. Chatting with an agent
(chat, ask) does respect it: with -O acme the usage is billed to the organization wallet and its model
policy applies.
Create and edit#
$ ipx agent create docs-bot --name "Docs bot" -m lightning \
--system "Answer from the docs." --visibility public --registry published \
--mcp docs=https://mcp.example.com/mcp
$ ipx agent update docs-bot --system-file new-prompt.md # only the options you pass change
$ ipx agent get acme-bot/docs-bot # owners/editors also see the prompt and members
$ ipx agent delete acme-bot/docs-bot
Option |
Meaning |
|---|---|
|
Display name (required on create) and a short description (500 characters at most). |
|
Backing catalog model (required on create). Agents cannot be backed by another agent or a Featherless deployment, and MCP servers need a model that supports tool calling. |
|
System prompt (32,000 characters at most). |
|
Private agents are only callable by you and invited members; public agents by any signed-in user. |
|
Public agents can be listed in |
|
Who pays for each call. |
|
Attach MCP servers. |
Discovering agents#
ipx agent list shows agents you own or were invited to (--mine for only yours). ipx agent registry -q TEXT
searches the public registry.
Settings#
$ ipx agent settings show
$ ipx agent settings set --include-in-models # also list agents in /v1/models for OpenAI-style tooling