Skip to main content
Agent skills provide LLMs and coding assistants with curated instructions, guidelines, and context-specific rules for executing tasks. MCP Assistant publishes a pre-configured skill, mcp-assistant, which instructs coding agents on how and when to use remote MCP servers, when to search for tools, and how to chain calls inside the CodeMode sandbox.

​
Installing the Skill

Install the MCP Assistant skill into your local workspace using the skills CLI: 
npx skills add zonlabs/mcp-ts --skill mcp-assistant
This downloads and configures the mcp-assistant skill folder in your workspace under .agents/skills/mcp-assistant/.

​
Updating the Skill

To update the skill to the latest version, run the update command: 
npx skills update mcp-assistant
If you are developing or testing local changes to the skill files in this repository, you can sync your local workspace files directly: 
npx skills add ./ --skill mcp-assistant

​
The mcp-assistant Skill

​
Core Principle

Keep MCP tools discoverable, not always loaded. Avoid loading every downstream MCP tool directly into the agent context, which can bloat the context window and cause tool-calling confusion. Instead, search for the relevant capability, inspect only the needed schema, then execute the selected tool or workflow through the sandboxed runner.

​
Available Tools

ToolCategoryPurpose
search_mcp_toolsSearchSearch connected MCP servers for candidate tools using natural language or exact names.
get_mcp_tool_schemaReadInspect the exact parameters and output shape for a selected tool.
codemode_runGeneralRun a TypeScript/JavaScript script inside a secure sandbox to execute tool chains.
When a task requires external tools or services connected to MCP Assistant, the agent follows this default flow:
  1. Identify the need: Determine if the goal requires an external MCP integration (e.g., GitHub, Notion, Supabase, Slack).
  2. Search tools: Call search_mcp_tools with a goal-oriented query (e.g. "find GitHub issues by label").
  3. Select candidates: Choose the most relevant tools from the search results.
  4. Inspect schema: Call get_mcp_tool_schema to retrieve the parameters and constraints for the chosen tools.
  5. Execute: Run the code/script using codemode_run to execute the tools and transform the output in the sandbox.

​
Safety & Approvals

Before executing destructive, public, or outbound communication actions, the agent must ask for user confirmation.
  • Needs Approval: Sending messages (Slack, Gmail, Discord), mutating production database records, closing issues, merging pull requests, or running expensive workflows.
  • Safe to run: Read-only queries, searching, listing, and local summarization.