Agents

An agent is a Claude Code agent file: a Markdown file whose front matter names it and describes it and whose body is its standing instructions. Localix reads the same folders Claude Code does, so an agent you write for one works in the other. There is no second agent format and no Agents page in Settings: agents are Store items, found and edited in the Store.

Create specialists from a conversation and manage them with /agents; see Create and manage specialists.

Where agents live

A file needs a name and a description. As in Claude Code, a name is valid unless it starts with -, contains : or is longer than 256 characters; such a file is skipped. Localix identifies an agent by that name, as written and with its case, so Code Reviewer and reviewer.v2 load under those names. A plugin's agent is plugin:name, or its bare name when one plugin alone has it.

description is the only text that says when to delegate: an agent has no when_to_use. Of the skills list the first 32 distinct names preload.

When two agents share a name, Claude Code's precedence decides which one runs: managed, then the agents mods register (Claude Code's --agents source), then the project (the closest folder wins), then yours, then plugins. A file with the name of a built-in type replaces it. Folders are watched: an agent you add or edit in an editor appears in the Store within a moment, and a file you delete leaves every list at once.

Localix keeps in its own database only what a file cannot hold: the agent's access rules, a stable identity that survives a rename, and chat history. The identity is why a renamed or moved file keeps its place in a conversation and its permission choices. A file that disappears is treated as missing for a grace period (10 minutes for your own folder, 7 days for a project) before its rules and conversation links are cleaned up; a disabled plugin or an unreachable project folder is never a deletion. Deleting an agent in the Store or with /agents delete cleans up at once. Messages it wrote stay, under its last name.

What your own agent files can set

Some front matter is honoured only in a file you or your organisation own (yours, the project's or a managed one) and in an agent a mod registers. A plugin's agent loads without it, as in Claude Code, because a plugin is code you did not write.

Project agents' hooks and inline mcpServers wait until Claude Code trusts the project folder, as a project's own settings do. A fork is its delegator rather than the agent, so it ignores all of these and keeps the delegator's hooks, servers and mode.

An inline MCP server connects when the child starts and disconnects when its run ends. A name must refer to a server that is already configured here; the child shares that server's connection, and the organisation's allow and deny lists apply to both kinds. A tools list that does not name a server's tools still withholds them.

A child runs in its conversation's permission mode. Only a conversation in Default lets the agent's own permissionMode apply, and only default and plan have a Localix mode: acceptEdits, auto, dontAsk and bypassPermissions change nothing. A conversation in Plan, Bypass or Autonomous keeps its mode in every child, whatever the file says.

Two more fields apply to every agent, a plugin's included. In disallowedTools, an entry removes the tool, with or without a specifier: Bash(rm *) removes Bash from the agent. In tools, Agent(worker, researcher) is an allowlist: an agent that runs as the main thread can delegate only to the types it names, and a call with no agent counts as general-purpose, so it is refused unless listed. A child cannot delegate at all, so the list has no other use.

The agent's page

Open Store → Agents, or follow an @name link in a chat. The page has these tabs:

• Overview: description, where the file is, how many chats use it, and the actions that fit: Edit, Duplicate, Copy to Yours, Move to a project or to Yours, Delete.
• Settings: the definition, written back to the file. Name, description, instructions, model, effort, turn limit and preloaded skills are fields; front matter Localix does not edit (tools, hooks, mcpServers and unknown keys) is kept as it is. A save that would overwrite a change made in an editor in the meantime is refused, and the page asks you to reload.
• Access: per-agent switches for tools, skills, MCP servers, plugins and mini apps. A switch is a rule for this agent only. Use Open on a row to go to that item's own page.

Agents from a plugin or a managed folder are read-only: Copy to Yours makes an editable copy with a fresh identity and no inherited permissions beyond the Localix default.

Localix, and the built-in types

localix is the main loop, not a file. It coordinates the requested outcome and handles straightforward work directly, and it cannot be deleted or disabled. Every conversation can fall back to it. To change how it behaves, edit CLAUDE.md or an output style, or set "agent" in your Claude Code settings.

general-purpose (also claude), Explore, Plan, statusline-setup and claude-code-guide are Claude Code's built-in subagent types and behave as in Claude Code. Explore and Plan answer once: a message to one is refused. Loops is not available to a subagent. See Claude Code compatibility.

You are not an agent. There is no @human record: describe yourself and your preferences in ~/.claude/CLAUDE.md or the project's CLAUDE.md. Agents never address you with an @ mention; to get your answer, a decision or an approval they use the AskUserQuestion tool (HITL); otherwise they report and stop, and the reply returns to you. A task-board mini app may still assign a card to you.

Starter agents

A starter team of ordinary plugin agents is in the Localix marketplace and Recommended in the Store. Nothing is added to your folders until you install it, and uninstalling removes them. Use Copy to Yours to change one. Pick one when you start a conversation, or name one as a subagent type.

Default tool candidates

Localix has an open candidate policy, so newly installed tools can join its runs. A candidate still needs a compatible run shape, host port and model. A child cannot acquire a tool that requires direct human interaction merely by allowing it in policy.

An agent-authored specialist receives a copy of its creator's saved permission rules. An agent made in the Store starts from the Localix default; narrow it on its Access tab and in Permissions. Definitions and priorities never grant permission.

Model, limits and context

In a new chat, the model route resolves from the choice you make in the composer, the model and effort you chose for the agent (its page in the Store), recent choices, the project and the primary route. A subagent runs on the model and effort you chose for it. Where you chose no model, it runs on the Agent call's model, its file's model or CLAUDE_CODE_SUBAGENT_MODEL, in Claude Code's order, else on its parent's route; its effort is otherwise its parent's (see Subagents).

The choice is kept in your Localix settings, never in the agent's file. The file's own effort is not read, and Localix never writes your choice into the file. A localix: key in an agent file is not read.

Turn limit (maxTurns) can lower a run's limit. Blank inherits the run limit, whose built-in default is 1,000 tool-loop turns; user settings and delegated task caps can lower it.

In the composer, click the turn-limit control to cycle through 1 → 5 → 20 → 100 → ∞ → 1. The tooltip shows the next value. ∞ uses the built-in 1,000-turn ceiling; the agent normally stops when its work is finished.

Skills (skills, at most 32) are preloaded for this agent in ordinary and delegated turns. Session-pinned skills are added, duplicates are removed, and denied skills remain unavailable. Agents do not inherit one another's skill lists. Loading a skill never grants denied tools or silences outside-workspace and dangerous-command confirmations.

Every agent in a chat shares the repository's memory folder. Claude Code's memory: key gives an agent run as a subagent a memory folder of its own, named by its type (a plugin's plugin:name as plugin-name); the user scope's is below CLAUDE_CONFIG_DIR when it is set, and with auto memory off the key does nothing. See stored context.

Agents delegate self-contained tasks to child sessions, run by themselves or by an agent file named in the task. An agent file is persistent; it is not a child run.

See Forks, Capabilities and Permission modes for the execution and policy contracts.