The Agent server is the management surface for 

 — the AI agents that read your documents and write into them. Through this server an application can list Charms, inspect their configuration, create new ones, edit their System Prompts, rename them, change models, copy them and delete them.

 a Charm (i.e. execute a completion), use the 

 servers. The Agent server is configuration, not execution.

 (the Charm Instructions document) and an 

 (model, temperature, tools, MCP servers, etc.). The two are bound permanently — a Charm's System Prompt is the 

 document it was created against and that binding doesn't move.

Multiple Charms can share the same System Prompt. Editing that document changes the behavior of every Charm bound to it. To diverge, copy the Charm document — the copy is a new Charm pointing at the same instructions, with its own independent configuration.

 accepts RMD. Omitted presentation is retained on

unchanged owners. When changing a styled owner’s content, retain its presentation

attributes explicitly, including lists, quotes and cells; ambiguous replacements

 to obtain the current source. The Charm Instructions

editor can remove styled blocks or replace them entirely. Source syntax is not

inferred from the instructions' text, and there is no 

Three paths, depending on what you start with:

The System Prompt binding is permanent. To "change" the instructions for an existing Charm, either edit the document with 

 (which affects every Charm bound to it) or create a new Charm.

 updates the model, temperature, tools, MCP servers and other LLM settings — and optionally renames the Charm. Before calling it, use 

 to learn what fields are valid for the provider and model you're targeting. The response includes:

 controls how Richtext Reference Material reaches the model. Use 

 to preserve message roles and tool rounds, or 

{ "type": "serialized", "format": "Markdown" }

 to include one serialized User message. The legacy 

This is how you avoid building an integration that pins a model name, then breaks when that model's capability surface shifts.

 — Execute completions with a Charm against a Chat document.

 — Ask a Charm a question inside a Workspace document.

List, create, configure, and manage Charms — everything except running them.

Tool	What It Does
`list_agents`	List the user's Charms
`list_agent_instructions`	List the available Charm Instructions (System Prompt documents)
`get_agent`	Full details for a Charm — config, model, tools, etc.
`get_agent_instructions`	The Charm's System Prompt content
`set_agent_instructions`	Overwrite the System Prompt content
`describe_model_config`	What configuration fields are valid for a provider/model combination
`create_agent`	Create a new Charm (from URI, inline text, or empty)
`config_agent`	Update an existing Charm's LLM configuration or name
`copy_agent`	Copy a Charm — same instructions, independent config
`delete_agent`	Hard-delete a Charm — not recoverable

Agent MCP Server

The Mental Model

Tools

Creating a Charm

Configuring a Charm

Patterns

Next Steps