Skip to content

Create and manage specialists

Agents are Claude Code agent files (see Agents), managed in the Store. A conversation has one agent, fixed when it is created, and a run is one piece of work. Creating an agent does not start a model run.

You are not an agent: there is no @human record, and describing yourself belongs in CLAUDE.md.

Create in the Store

Open Store → Agents and choose Add agent. Give it a name and a description, and choose where to keep it: yours (~/.claude/agents/) or the open project (<project>/.claude/agents/). A name already taken in that place is refused. Localix writes the file with a placeholder body and opens the agent's page on its Settings tab, where you write the instructions.

Create from a conversation

Ask your assistant to create a specialist. With the Store tool available, it can look at what is installed (search with kind: "agent") and, once you approve, create a missing specialization with create, giving a name, a description and the standing instructions. Current task details belong in the task you give it rather than in the reusable definition.

Every create shows you a card with the definition and waits for your answer; nothing is written before you approve. A run nobody can answer, such as an automation, refuses it. The scope is user (your agents, in every project) or project (the open project's .claude/agents/ folder); creation cannot select a different project. Your own agents and plugin agents such as the starter team are listed in every project too, unless a project agent uses the same name. The server issues a stable ID and leaves model preferences, memory settings, skill preloads and turn limits at their existing defaults.

A created specialist starts from the Localix default policy, as with the Store's Add agent form. It does not copy its creator's permission rules; edit them on its Access tab when needed.

A created specialist is available to start a new conversation with, or to name as a subagent type in a task (see Subagents). Child and automation runs do not receive the Store tool.

The localix agent receives team-building instructions. It handles small tasks directly; when distinct specialties are useful, it reuses existing definitions and creates the missing roles one Store call at a time, each one approved by you.

User commands

Agents are managed in the Store: the Agents shelf adds, edits, copies, enables and removes them, and an agent's own page has its Access tab. The /agents command only lists and points there.

CommandResult
/agents listList the agents of this chat's project by name, including disabled ones.
/agentsLink to the Agents shelf of the Store, where Add agent is.

An agent is identified by its Claude Code name; there is no @handle form. To make an agent in your own words, ask the chat's agent: it picks the name, description and instructions and saves the specialist with the Store tool after you approve. It does not start the new agent.

Managed, plugin and built-in agents cannot be edited or deleted, exactly as in the Store. Deleting an agent checks whether any conversation still uses it.

Availability and retries

An agent with disabled availability or a denied chat invocation cannot start a conversation.

Repeated delivery of the same creation call returns the saved identity, even after a rename. Reusing that operation for different instructions fails. A different creation with an occupied name fails without overwriting the existing agent.

Localix is the only built-in agent. Customized policies remain unchanged; enable the Store tool on the agent's Access tab when desired. Its reads (search, get) can be granted independently of its changes, and a rule can name one package, such as install:plugin:review@market:user.