Plugins and skills

Plugins in the left sidebar is one combined entry point: add capabilities to your workspace and use them across your agents. It has only two tabs — MCP and Skills — and opens on Skills.

The difference in one line: an MCP is a running process that exposes tools to agents; a skill is a Markdown file that tells an agent how to do something.

Skills

One page in two parts: Installed on top, Marketplace below, both as cards. "Installed" here means enabled for at least one agent — a skill that ships with Croli but that you have not switched on counts as part of the marketplace half, next to the skills.sh entries.

The switch on a card applies to every skills-capable agent at once; there is no "enable it for just this one agent". Cards in the marketplace half carry a + instead of a switch, and pressing it is what enables them.

Each card has two tags: its category, and its source — Experts / Research / Office / Custom. It may pick up Modified, Needs API key or Needs setup as well. An Office skill that is not ready has its switch greyed out, and its detail view carries an amber line — "Install or sync the office tools before enabling this skill" — with the Set up button right next to it.

The toolbar holds a search box, a category filter (17 categories) and Create, with Import / Import from agent / Refresh behind the ⋯. The search box filters the installed half as you type and only queries skills.sh when you press Enter (the line under the marketplace heading says as much: "Featured on skills.sh — press Enter to search"), matching a skill's name, ID and description. Import from agent pulls an agent's global skills into your custom skill library.

Click a card to open its detail, which is a package reader: a Files rail on the left, the selected file's content on the right. A custom skill's detail adds a ⋯ menu — Edit / Duplicate / Delete. A skills.sh card instead offers Install skill, View on skills.sh and its install count.

Where skill files land

Enabling a skill for an agent means, on disk, writing it into that CLI's own skills directory — and each CLI decides where that is. "Global" is shared across projects; "Project" is relative to the current workspace root by default — except for DeepSeek Harness, which looks for the nearest ancestor directory containing .git (falling back to the current directory if it reaches the filesystem root without finding one):

~/.agents/skills and the project's .agents/skills are a shared library: a skill you put there is visible to every agent that reads that directory.

MCP

One page in two parts again: Installed on top, MCP Marketplace below. The toolbar holds the marketplace source selector, a search box, Refresh and New MCP. The search box filters the installed half as you type, and queries the selected marketplace on Enter.

New MCP

Three fields:

• Server ID — an identifier you pick.
• Which agents it applies to — see the hard limit below.
• MCP Config (JSON) — the default draft is {"type":"stdio","command":"","args":[]}. The supported types are stdio / http (aliases streamable-http, streamableHttp) / sse. env only applies to stdio; a remote MCP's auth token goes in headers.

One hard limit: only four agents

"Which agents it applies to" offers only four options for an MCP server:

Claude Code, Codex, Grok, DeepSeek Harness.

Those four are the only assignable targets. The installed list, though, merges the config files of fifteen different CLIs, so you may well see servers belonging to the likes of Cursor or OpenCode here — their assignment cannot be changed, and to touch them you have to edit that CLI's own config file.

When an agent's MCP config file cannot be read, an amber warning appears: that agent's servers are hidden, and saving is blocked too — fix or delete the file first, then press Refresh.

The installed half

Each card's ⋯ menu holds Edit / Uninstall, and the card carries an auth-status tag. A server that needs OAuth gets an extra Accounts section in its edit sheet (connect / reconnect / disconnect; the credentials are kept by the CLI itself, never by Croli); a plain stdio server has no such section. Right after an install you also get one "Connect your account" prompt.

MCP marketplace

Two fixed sources, which you cannot add to or remove:

Click a card to open its detail sheet, where you pick the Default Install Protocol, expand Advanced settings to edit the install JSON, then Install. If you would rather skip the detail, the + on the right of the card goes straight to the install confirmation. Both install and uninstall have a confirmation dialog.