See which skills, MCP servers and hooks Claude Code loads, and turn them off

You added a GitHub MCP server with claude mcp add under your work account. A week later you open the same repo under your personal account, and it isn't there.

Updated October 15, 2026

That server lives in the work account's own .claude.json. To see what each account loads, you'd read one of those files per account, then the project's settings and every plugin.

Skills drift too. You copy one into ~/.agents/skills so Codex has it, and now there are two. And a skill you last used four months ago still loads in every project.

# the dashboard: choose Extensions
clausona

# which accounts have github, and whether it's on here
clausona mcp show github

# Claude skills not used in 90 days
clausona skills ls --scope unused --tool claude

• What loads in this project
• Which account loads which MCP server
• Turn one off, here or everywhere
• Clear out skills you don't use
• Undo a change
• From the command line, or from an agent
• Questions

What loads in this project

Run clausona and choose Extensions. The screen looks from the project you started it in: the git repository around the current folder, or the folder itself outside one.

The first line on the left is that project, with a count of everything that loads there. The table on the right groups those rows by where they come from. The project's own come first, then the other places, in the order of the list on the left.

Claude and Codex are tabs, and tab switches between them. 1, 2 and 3 switch between skills, MCP servers and hooks.

Under the first line, each place has a line of its own. Open one and you also see what doesn't load, such as a skill a same-name copy hides, or one you turned off here.

enter on a row opens its details. They show its files and its state in each account, and for a skill, how often it's been used.

To look from another project, press p. It lists the folders Claude Code or Codex recorded in any of your accounts, and typing narrows the list.

Which account loads which MCP server

MCP servers are where accounts differ most.

Plugins are shared, so a server that comes with a plugin is there under every account. A server added with claude mcp add is stored in one account's .claude.json, and only that account starts it.

On Claude's MCP tab, m opens a matrix. Each server gets a row, and each account a column, with one more column for Codex. A cell says whether that server starts for that account in this project.

space on a cell turns that one account's server off or on here, and the other accounts stay as they were.

A server from the project's .mcp.json is the exception. Claude Code switches it for the whole project, so its cell changes it for every account.

From the command line:

clausona mcp show github

It says which accounts have the server, and whether it's on here. Secret values in its config are never shown.

Turn one off, here or everywhere

Select a row and press space to turn it off in this project. g turns it off in every project.

Before anything changes, a dialog lists each file and what will change in it, and where the backup goes. y goes ahead, and n leaves everything as it was.

clausona writes the settings Claude Code and Codex already read.

A Claude skill turned off here is set to "off" under skillOverrides, in the project's .claude/settings.local.json. A Claude MCP server from an account's .claude.json, or one a plugin brings, turned off here goes into disabledMcpServers in that account's .claude.json, under its entry for the project. That's the list /mcp disable writes. A server from the project's .mcp.json goes into disabledMcpjsonServers in .claude/settings.local.json instead, for every account at once.

For Codex, a server is turned off with enabled = false. Here, that goes in the project's .codex/config.toml. Everywhere, which works for your own servers, it goes in ~/.codex/config.toml.

A few things switch differently:

• A hook has no switch for one project, in Claude Code or in Codex. Off takes it out of its settings file in every project, and clausona keeps it so that g can put it back.
• What a plugin brings is switched with the plugin itself. A plugin's MCP server is the exception: it also has a switch of its own, per account, in each project.
• Codex's plugins are listed, but you switch them in Codex, with /plugins.
• A server from the project's .mcp.json can only be switched in one project at a time. Claude Code has no switch for it everywhere.

When a row can't change, nothing opens, and the status line says why.

Clear out skills you don't use

Skills pile up. One you tried once and forgot about keeps loading until you remove it.

The Not used in 90 days line lists the Claude skills no account has used in 90 days. Skills never used at all show up there once they're more than 14 days old.

Use is added up over all your accounts. A skill your work account uses every day isn't on the list, even if your personal account has never touched it.

clausona skills ls --scope unused --tool claude

d deletes the selected skill, or every row you marked with x. The folder moves into a backup instead of being removed. A skill that's only a link loses the link, and the folder it points to stays.

A skill git tracks belongs to the repo, so deleting it changes the repo. The screen warns you in the dialog and offers to turn it off here instead. The CLI refuses unless you add --tracked.

A Cloud skill comes back from claude.ai, so turn it off instead of deleting it. A plugin is uninstalled with /plugin in Claude Code.

Undo a change

Every change starts with a backup. clausona copies each file into ~/.clausona/backups/extensions/ before it writes, and a deleted folder moves there. The last 50 changes are kept.

u undoes the newest change, after asking. From the command line, each kind has its own:

clausona skills undo
clausona mcp undo
clausona hooks undo

Run it again to go one change further back.

Undo puts back only what still holds what the change wrote. If you or Claude Code changed that since, undo leaves the file alone and says so. In a JSON file such as .claude.json, which Claude Code rewrites all the time, only the keys the change wrote count.

From the command line, or from an agent

Everything on the screen has a command. clausona skills, clausona mcp and clausona hooks each take ls, show, off, on, rm and undo.

clausona skills ls --scope project        # the skills this project defines
clausona mcp off github --account work    # off here, for work only
clausona skills rm old-one --dry-run      # what deleting it would change

A change prints its plan and asks first. --dry-run stops at the plan, and --yes skips the question.

A coding agent has no terminal to answer on, so it needs --yes. With --json, every command prints one JSON object, plans and errors included.

The Extensions docs list every scope and option, with recipes for agents.

Questions

Why is my MCP server missing under my other Claude Code account?

A server added with claude mcp add is stored in the .claude.json of the account that added it, and clausona doesn't share that file between accounts. The matrix (m on the MCP tab) shows which accounts have it. Add it under the other account too, or put it in the project's .mcp.json, which every account reads.

Does turning a skill off delete it?

No. Off writes a setting, and the skill's folder stays where it is. For Claude Code that's skillOverrides in a settings file, and for Codex an entry in config.toml. Press the same key again to turn it back on, or run clausona skills on, adding --everywhere if it was off everywhere. Deleting is a separate step, and it moves the folder into a backup.

Can I turn off an MCP server in one project only?

Yes. A server from an account's .claude.json, or one a plugin brings, is switched per account. clausona adds its name to disabledMcpServers in that account's entry for the project, the list /mcp disable writes, and only accounts that have opened the project have that entry. A server from the project's .mcp.json is switched for every account at once: off adds its name to disabledMcpjsonServers and takes it out of enabledMcpjsonServers, in the project's .claude/settings.local.json. For Codex, off writes enabled = false in the project's .codex/config.toml, in a project Codex trusts.

Can I turn off a hook in one project only?

No. Neither Claude Code nor Codex has a switch for one hook in one project, so a hook goes off everywhere. clausona takes it out of its settings file and keeps it, and turning it back on puts it back.

Will it change files in my repo?

Turning a Claude skill or a .mcp.json server off here writes the project's .claude/settings.local.json. Claude Code keeps that file out of git with a line in your global git ignore file, and clausona adds the same line when it creates the file and git doesn't ignore it yet. A change to a file git tracks, such as a skill folder in the repo or the project's .codex/config.toml, needs --tracked in the CLI, or your yes in the screen's dialog.

Does it show my API keys or tokens?

No. An MCP server's env and header values are never shown, only their names. In a command or a URL, a value that looks like a secret reads <hidden>. Nothing leaves the machine. The backups do hold whole entries, secrets included, so only you can read their folders.

Does it work with Codex?

Yes, on the Codex tab, with the skills, MCP servers and hooks Codex loads in the project. Codex reads a project's .codex folder only when it trusts the project, and clausona follows that. Codex's plugins are listed, but you switch them in Codex, with /plugins.