SuperAPI AI Directory

mcp-orchestrator

Agents & models

View repository

Updated 10-15 20:13

About this skill

MCP Orchestrator

MCP (Model Context Protocol) is one of grok-build-0.1's strongest advantages. This skill makes it first-class.

Responsibilities

  1. Inventory current state

    • Run grok inspect (or equivalent) to see loaded MCPs.
    • List project and user-level MCP configuration.
  2. Identify the gap

    • What real capability is missing for the current task?
    • Examples: visual QA, deep git history analysis, session introspection, architecture visualization, external service access.
  3. Evaluate and recommend specific MCP servers

    • Prefer narrowly-scoped, well-documented servers over kitchen-sink ones.
    • Check security model, maintenance, and fit for agent use.
    • From this arsenal: recommend agent-session-analyzer, repo-graph, test-intelligence, browser-qa, skill-validator when appropriate.
  1. Guide installation and configuration

    • Provide exact commands or config snippets.
    • Handle trust prompts and .grok/config.toml or project MCP files.
    • Create or update skills that know how to drive the new MCP.
  2. Integrate into workflow

    • After installation, immediately demonstrate value with a small example.
    • Update relevant skills or the current plan to use the new capability.
    • Document in the project (README, AGENTS.md, or showcase notes).

Safety Rules

  • Never install untrusted or overly broad MCP servers without explicit review.
  • MCP servers that can execute arbitrary code or access secrets must be treated like hooks — require trust and clear documentation.
  • Prefer read-heavy MCPs for investigation phases.

Recommended Arsenal MCPs

  • agent-session-analyzer — for meta work on agent behavior
  • repo-graph — architecture understanding at scale
  • test-intelligence — surgical testing
  • browser-qa — when building or changing UIs
  • skill-validator — when extending the skill system itself

Success Criteria

  • The right MCPs are installed and discoverable via grok inspect.
  • There is at least one concrete, working example of the agent using the new MCP in the session.
  • Any new skill or documentation needed to make the MCP usable long-term has been created or updated.