Resources

Overview
What it is	The shared operations for all seven resource types
CLI	agt provider, model, prompt, skill, mcp, agent, harness resource
Agent tools	resource_* (needs resource_management)
Web console	Resources (资源库), Agent team (Agent 团队)
Crates	agt-core, agt-storage, agt-service

What it does

Resources are the files that describe agents. There are seven types, each with its own page:

Type	Describes	Page
Provider	A model API endpoint and its key	Provider
Model	A model on a provider and how to call it	Model
Prompt	A role description	Prompt
Skill	A reusable instruction package	Skill
MCP	An MCP server that provides tools	MCP
Harness	An external engine (Codex) and its project	Harness
Agent	A model or harness plus prompt, skills, MCP servers and capabilities	Agent

All seven share the same naming, categories, deletion rules and agent tools, described here. People and agents edit resources through the same code path, with the same validation and locks.

How to use it

agt model list                         # list
agt model info fast                    # show one
agt model category create experiments  # group resources
agt model create cheap --category experiments --provider main --model gpt-5-nano
agt model delete experiments/cheap --dry-run

An agent with resource_management does the same through tools:

{"kind": "model", "reference": "experiments/cheap"}

Common parameters

Parameter	Meaning
name or category/name	How a resource is referenced. A bare name means the default category.
--category <c>	Category of the resource being created or looked up
--recursive	When deleting, also delete what depends on the resource
--dry-run	When deleting, show the plan only

All parameters

Shared CLI commands

Every resource type supports these subcommands (for harnesses: agt harness resource ...). Aliases: add = create, ls = list, show/get = info, edit = update, remove = delete.

Command	Options	Meaning
list	--category <c>	List resources, optionally in one category
info <ref>	--category <c>	Show one resource (secrets redacted)
delete <ref>	--category <c> --recursive --dry-run	Delete a resource
category create <name>		Create a category
category list		List categories
category delete <name>		Delete an empty category other than default

create, update and type-specific commands are on each type's page.

Deletion rules

• Deleting a resource that others still reference is refused, with the list of dependents.
• --recursive deletes the dependents too: deleting a provider deletes its models and the agents using them; deleting a prompt, skill, MCP server, harness or model deletes the agents using it. Resources an agent uses are never deleted because the agent is deleted.
• An agent that is still bound must be unbound first (see Agent).
• default/default (prompt) cannot be deleted.
• Before deleting, agt checks all seven types for broken configuration and dangling references, and refuses if anything is inconsistent.
• Deleted directories are moved to <home>/.trash/. In a terminal you are asked to confirm; elsewhere pass --yes.

Agent tools

Tool	Parameters	Meaning
resource_list	kind	List resources of one type (secrets redacted)
resource_read	kind, reference	Read one resource, including its revision
resource_write	kind, reference, create, data, expected_revision?, add?, remove?, clear?	Create (create: true) or update a resource
resource_delete	kind, reference, recursive?, confirm?	Delete in two steps (below)
resource_category_list	kind	List categories
resource_category_create	kind, category	Create a category
resource_category_delete	kind, category	Delete an empty category other than default

kind is one of provider, model, prompt, skill, mcp, agent, harness.

resource_write details. The fields of data depend on the type and are listed on each type's page. On update:

• omitted fields stay unchanged;
• expected_revision (from resource_read) makes the update fail if someone changed the resource in the meantime;
• add / remove add or remove items of an agent's skills, mcps and builtin_tools;
• clear resets optional fields: an agent's description, skills, mcps, builtin_tools; a model's max_tokens, reasoning_effort; a harness's model, reasoning_effort.

Do not combine replacing a list, clear, and add/remove of the same item in one call. add, remove and clear are not allowed with create: true.

resource_delete details. Call it without confirm to get the deletion plan. Then call it again with confirm set to the exact, ordered list of items in the plan. An empty list never confirms a non-empty plan.

Related modules

• Workspace: where resource files live
• Agent: how resources come together into an agent