Use skills and MCP servers
Bring in your skills and MCP servers, pick them with $, and turn servers on or off with /mcp.
Skills are instructions for a kind of work, such as a review checklist or a deploy process. MCP servers give the agent tools and data from other services, such as Linear or Figma. Jinion brings in both from where you already keep them.
Where they come from
| Source | Comes in as |
|---|---|
~/.claude/skills, ~/.agents/skills | Your skills, such as $design |
.claude/skills, .agents/skills in the project | Project skills; they win over a user skill with the same name |
| Plugins you turned on in Claude Code | Their skills, commands, agents and MCP servers; their hooks stay off |
~/.jinion/mcp.json, ~/.claude.json | MCP servers |
The project's .mcp.json | MCP servers that stay off until you turn them on |
| Your Claude account | Its claude.ai connectors |
Use a skill
Type $ and pick one. A skill can go anywhere in the message, and you can pick several:
review the open PR $reviewA message that starts with a skill runs it, with the rest as its input. The agent also loads a skill on its own when
one fits the task, and the conversation shows Skill design.
Typing /design out of habit puts $design back in the prompt, and Jinion says
Skills go after $ now, anywhere in the message: $design. Press enter to send it.
Use an MCP prompt
MCP servers can offer prompts, listed under $ with the server's name. The rest of the message goes in as the
prompt's input:
$pr-summary 128Turn servers on or off
/mcp lists every server with its state and where it comes from. The state is connected with its number of tools,
connecting…, not signed in, failed to connect, or off. The one you're on shows its URL or command, its error and
its tools.
| Key | Does |
|---|---|
space | Turns the server on or off |
enter | Saves; applies from the next turn |
esc | Closes without saving |
After enter, Jinion says what changed, such as Turned on linear. This applies from the next turn. The conversation
carries on.
A server that shows not signed in needs a login: ask the agent to authenticate it, or, for a claude.ai connector,
connect it in your claude.ai settings.
Servers in a project's .mcp.json stay off until you turn them on, since anyone with commit access can change that
file. They show off until you turn it on (from .mcp.json).
Tool calls and context
MCP tools are listed to the model by name only, and loaded when used. Many servers cost little context this way. Each call asks for permission like any other tool, with "don't ask again" for that tool in the project.
In the conversation, a call shows as server:tool (MCP) with its arguments, and the first lines of what came back
below it. Click it, or press ctrl+o, to see all of it.