Agents

An agent is a reusable assistant definition: a Markdown body of instructions plus structured frontmatter that controls which tools and model it may use. You author it once in cnfgr, and it’s written into each connected tool’s native format.

Anatomy of an agent

cnfgr stores agents in Claude Code’s shape (the richest of the supported tools) and derives other tools’ formats at write time.

Field	Meaning
`name`	Identifier used to reference the agent.
`description`	What the agent is for and when to use it.
`tools`	Tools the agent is allowed to use.
`disallowedTools`	Tools explicitly denied.
`model`	Preferred model.
`permissionMode`	How the agent asks for permission to act.
`mcpServers`	MCP servers the agent may use.
`skills`	Skills available to the agent.
`maxTurns`	Cap on the number of turns.
body	The Markdown instructions for the agent.

Not every field maps cleanly onto every tool. See the tool support matrix for where fidelity is reduced.

Editing an agent

Open the Agents tab and select an agent (or create a new one). The editor has:

Frontmatter fields — name, description, tools, model, and the rest of the structured options above. Field labels carry inline help so you don’t have to remember each one’s purpose.
Body editor — write the agent’s instructions in Markdown. You can switch between a raw Markdown view and a Rich Text view of the same content.

Saving an edit records it in the canonical model and projects it into every connected tool, then syncs it to your other devices.

Global vs project agents

An agent can be global (available everywhere) or scoped to a project. When a project agent shares a name with a global one, the project version takes precedence and the global one is marked overridden by project. See Projects & identity.