Agent extensions

Before a turn starts, an agent loads more than your prompt: instruction files such as AGENTS.md and CLAUDE.md, rules, skills, commands, subagents, MCP servers, hooks and plugins. Settings ▸ Agent extensions gathers what Claude Code, Codex and Cursor will load into one list, says when each piece reaches the model, and estimates what the always-on part costs every turn.

The provider’s files stay the source of truth. Codz lists them without importing them, and it never opens a file that can hold credentials or commands, such as a settings, hooks or MCP file. It shows that such a file exists and leaves its values alone.

Open the inventory

Open Settings ▸ Agent extensions. The Inventory card sets what is listed:

• Scope follows the chat selected in the main window. Project extensions are looked for in that chat’s folder and in each folder above it, up to the root of its repository. With no chat selected, only user and managed extensions are listed. The scan runs again when you select another chat, and Refresh runs it on demand.
• Provider narrows the list to Claude, Codex or Cursor, or shows All. A skill in a shared .agents/skills folder appears once, with both providers named.
• Find searches names, kinds, scopes, providers and paths.

Four figures sit under the card: an estimate of the context loaded on every turn, the number of skills, the number of tools and automations (commands, subagents, MCP servers, hooks and plugins), and how many entries need review. With All providers, the first figure is labelled Max / turn est. and shows the largest single provider’s total, because no turn loads all three; for one provider it reads Always-on est.

Read the list

Entries are grouped by kind: Instructions, Skills, Rules, Commands, Subagents, MCP servers, Hooks, Plugins and Provider configuration. Each entry shows its name, its provider, its scope (Project, User, Managed or Installed), when it reaches the model, and its path. The arrow button reveals the file in Finder.

The “when” tag is one of: every turn, when matched, on demand, separate context, tool catalog, on lifecycle event, no prompt cost, component dependent, or provider managed. Text that reaches the model carries an estimate, such as “≈… every turn”. Estimates come from file size and are marked with ≈; they are not token counts.

An entry tagged Review needs a look, and says why:

• the file is large for its kind — over 32 KB for instructions, 512 KB for a skill or 64 KB for anything else — so its context cost is worth checking;
• a skill is missing its name or description frontmatter, or links to a folder with no SKILL.md;
• two skills for one provider share a name, so the provider’s precedence decides which one runs.

An entry tagged Protected is a file whose values Codz deliberately doesn’t read.

How context is used

Near its end, the page sums up how each kind reaches the model:

What Codz looks for

“Project” means the chat’s folder and the folders above it, up to the repository root.

Managed entries come from places the provider or an administrator controls: Claude Code’s CLAUDE.md in /Library/Application Support/ClaudeCode, and Codex skills in /etc/codex/skills and ~/.codex/skills/.system. Codz tags a Claude Code rule as every turn unless it names paths, and a Cursor rule only when it sets alwaysApply: true; the rest are tagged when matched. OpenCode’s extensions are not part of the inventory.

Manage integrations

To add or change plugins, plugin sources and MCP servers, click Manage under Manage integrations, or choose Chat ▸ Agent Integrations…. The Agent integrations window runs the provider’s own CLI commands; Codz fills in the form and shows the terminal.

1. Choose the provider and the type: Plugins, Sources or MCP servers.
2. Choose the account and the scope. For Claude Code and Codex, the account is CLI configuration or one of your additional logins. Claude Code offers the scopes Account, Project (shared) and Project (personal); the others change the account-wide configuration only.
3. Choose the operation and fill in the identifier or source it asks for.
4. Read the command and what it will change, shown before anything runs.
5. Click Run.

The command runs in the window’s terminal, where the provider can still ask you questions, and Interrupt stops it. The output stays in that window and isn’t added to a chat. While an operation that changes configuration runs, that provider’s chats wait; when it finishes, Codz restarts the provider’s idle agent processes so the next turn loads the new configuration. An operation started from the Manage button refreshes the inventory when it ends.

• Discover has two shortcuts for Claude Code: Official Claude catalog fills in Anthropic’s plugin source, and GitHub for Claude fills in the GitHub plugin. You still review and run them.
• Adding an MCP server takes an HTTPS address without credentials or a query; the provider handles signing in separately, with Authenticate where it offers one. OpenCode asks for the server’s details in the terminal instead.
• Skills aren’t installed here. A skill that comes in a plugin is managed through its plugin; a standalone skill is a file you edit or remove yourself, which the inventory reveals in Finder.
• Provider guide opens the provider’s own documentation for the chosen type.

Codz’s own additions

The inventory lists the providers’ files only. Codz adds a layer of its own to every prompt you send, on every provider: the Personality, Custom instructions and, when turned on, Memory from Settings ▸ Agent & composer. See Settings.

Provider guides

Each provider’s own documentation on customizing its agent is one Open guide away at the foot of the page:

• Codex — skills, AGENTS.md, MCP, rules, hooks and plugins: learn.chatgpt.com/docs/build-skills
• Claude — instructions, skills, subagents, hooks, MCP and plugins: code.claude.com/docs/en/features-overview
• Cursor — workspace and user customizations, including plugins and skills: cursor.com/changelog/customize

Troubleshooting

The list is empty. With no chat selected, only user and managed extensions are listed: select a chat to include its project. If a search or a provider filter is on, clear the search or choose All.

An operation says the installed CLI doesn’t support it. Before each run, Codz checks that the installed CLI has the command and options. “This installed CLI does not support the requested operation and options. Update it or open the provider guide.”

An operation asks you to wait. A command that changes a provider’s configuration waits until none of that provider’s chats is running a turn: “Wait for this provider’s active tasks or integration operation to finish, then run again.”

You suspect an extension is causing trouble in Claude Code. Turn on Safe mode under Settings ▸ Models & providers ▸ Claude. From the next turn, Claude Code runs with CLAUDE.md, skills, plugins, hooks, MCP servers and other customizations disabled. Turn it off again when you are done.