MCP Code Execution

MCP Code Execution Hub

Quick Start

This skill is an orchestration hub, not a CLI. It activates inside a Claude Code session when one of the trigger keywords below appears, or when invoked explicitly:

Skill(conserve:mcp-code-execution)

The hub then routes to the relevant sub-skill modules (mcp-subagents, mcp-patterns, mcp-validation) based on the detected workflow shape. There is no separate install step or CLI entry point.

When To Use

• Automatic: Keywords: code execution, MCP, tool chain, data pipeline, MECW
• Tool Chains: >3 tools chained sequentially
• Data Processing: Large datasets (>10k rows) or files (>50KB)
• Context Pressure: Current usage >25% of total window (proactive context management)

MCP Tool Search (Claude Code 2.1.7+): When MCP tool descriptions exceed 10% of context, tools are automatically deferred and discovered via MCPSearch instead of being loaded upfront. This reduces token overhead by ~85% but means tools must be discovered on-demand. Haiku models do not support tool search. Configure threshold with ENABLE_TOOL_SEARCH=auto:N where N is the percentage.

Subagent MCP Access Fix (Claude Code 2.1.30+): SDK-provided MCP tools are now properly synced to subagents. Prior to 2.1.30, subagents could not access SDK-provided MCP tools: workflows delegating MCP tool usage to subagents were silently broken. No workarounds needed on 2.1.30+.

Claude.ai MCP Connectors (Claude Code 2.1.46+): Users logged into Claude Code with a claude.ai account may have additional MCP tools auto-loaded from claude.ai/settings/connectors. These tools contribute to the tool search threshold count. If workflows unexpectedly trigger tool search or context inflation, check /mcp for claude.ai-sourced connectors. Known reliability issue: connectors can silently disappear (GitHub #21817).

MCP Prompt Cache Fix (Claude Code 2.1.70+): MCP servers with instructions connecting after the first turn no longer bust the prompt cache. Previously, a late-connecting MCP server would invalidate cached prompt prefixes, increasing token costs for the rest of the session. On 2.1.70+, prompt cache reuse is preserved regardless of when MCP servers connect.

ToolSearch Reliability Fix (Claude Code 2.1.70+): Empty model responses after ToolSearch are fixed. The server was rendering tool schemas with system-prompt-style tags that could confuse models into stopping early. ToolSearch-heavy workflows (many deferred MCP tools) are now more reliable.

When NOT To Use

• Simple tool calls that don't chain
• Context pressure is low and tools are fast

Core Hub Responsibilities

• Orchestrates MCP code execution workflow
• Routes to appropriate specialized modules
• Coordinates MECW compliance across submodules
• Manages token budget allocation for submodules

Required TodoWrite Items

1. mcp-code-execution:assess-workflow
2. mcp-code-execution:route-to-modules
3. mcp-code-execution:coordinate-mecw
4. mcp-code-execution:synthesize-results

Step 1 – Assess Workflow (mcp-code-execution:assess-workflow)

Workflow Classification

def classify_workflow_for_mecw(workflow):
    """Determine appropriate MCP modules and MECW strategy"""

    if has_tool_chains(workflow) and workflow.complexity == "high":
        return {
            "modules": ["mcp-subagents", "mcp-patterns"],
            "mecw_strategy": "aggressive",
            "token_budget": 600,
        }
    elif workflow.data_size > "10k_rows":
        return {
            "modules": ["mcp-patterns", "mcp-validation"],
            "mecw_strategy": "moderate",
            "token_budget": 400,
        }
    else:
        return {
            "modules": ["mcp-patterns"],
            "mecw_strategy": "conservative",
            "token_budget": 200,
        }

MECW Risk Assessment

Delegate to mcp-validation module for detailed risk analysis:

def delegate_mecw_assessment(workflow):
    return mcp_validation_assess_mecw_risk(
        workflow, hub_allocated_tokens=self.token_budget * 0.5
    )