/mcp-toolchain-manager

name: mcp-toolchain-manager
description: 'Detect frontend tool capabilities, report gaps, verify official setup sources, plan approval-gated MCP changes, or update project/mcp-profile.md. Prefer capabilities over provider names; never install or configure without approval.'
id: 'agents.skills.mcp-toolchain-manager.skill'
title: 'MCP Toolchain Manager'
doc_type: 'skill'
layer: 'skill'
status: 'active'
publishable: true
local_only: false
skill: 'mcp-toolchain-manager'
tags:
    - 'agents/skill-package'
    - 'agents/tooling'
    - 'workflow/mcp'
parent:
    - '[[AGENTS|Canonical Agent Policy]]'
related:
    - '[[common/tool-capability-model|Tool Capability Model]]'
    - '[[common/mcp-installation-policy|MCP Installation Policy]]'
    - '[[common/prompt-intent-routing-rules|Prompt Intent Routing Rules]]'
    - '[[skills/goal-planner/SKILL|Goal Planner]]'
    - '[[skills/execution-plan-manager/SKILL|Execution Plan Manager]]'
    - '[[skills/project-onboarding-adapter/SKILL|Project Onboarding Adapter]]'
    - '[[skills/project-context-adapter/SKILL|Project Context Adapter]]'
    - '[[skills/agent-rules-skill-author/SKILL|Agent Rules And Skill Author]]'
depends_on:
    - '[[AGENTS|Canonical Agent Policy]]'

MCP Toolchain Manager

Purpose

Detect, explain, validate, and record tool capabilities needed by WebDev Agent Kit skills without turning tool setup into automatic or unsafe behavior.

This skill exists to keep MCP usage deliberate. It identifies which capabilities are required for the current workflow, which providers are already available, which providers are missing, what each missing capability blocks, which fallback is allowed, and what official source must be verified before installation or configuration is proposed.

When To Use

Use this skill when:

• a selected skill declares capability dependencies in `tool-capabilities-manifest.json`;
• a task requires rendered UI verification, official documentation lookup, project context retrieval, visual references, GitHub metadata, or another external capability;
• the agent needs to audit available vs missing MCP servers, native tools, connectors, or fallbacks;
• the user asks what MCP servers or tools are required for the skill pack;
• the user asks to install, configure, validate, or troubleshoot MCP tools;
• `project/mcp-profile.md` must be created or refreshed;
• a missing tool limitation must be reported before fallback behavior.

Use this skill in `Standard Workflow` or `Deep Workflow` when tool capability affects the current slice.

It may also be used in `Lightweight Workflow` only when the user directly asks about MCP/tool availability or when a missing already-required capability prevents a narrow task from being verified.

When Not To Use

Do not use this skill just because a task exists.

Do not use this skill for:

• one small bug when no external tool is needed;
• one obvious type error;
• one small styling adjustment that does not require rendered verification;
• documentation-only edits unrelated to MCP/tooling;
• generic planning that does not depend on tool capabilities.

Do not install tools from this skill unless the user explicitly approved that installation in the current task context and the official source was verified.

Do not use this skill to implement frontend code, write tests, scaffold projects, create UI library setups, or access production systems.

Required Context

1. Read `AGENTS.md`. 2. Read `common/tool-capability-model.md`. 3. Read `common/mcp-installation-policy.md` when installation or configuration is in scope. 4. Read `common/prompt-intent-routing-rules.md` when workflow level is unclear. 5. Read `tool-capabilities-manifest.json` for selected skills and capabilities. 6. Read `common/codex-official-docs-policy.md` only when `openai_platform_docs` is active. 7. Read selected `agents/openai.yaml` files only when Codex UI or invocation metadata matters; never use them as capability declarations or availability evidence. 8. Read `project/mcp-profile.md` when it exists and the task needs durable tool state. 9. Read `project/verification-profile.md` only when verification commands influence tool needs. 10. Do not read all skills or all YAML files during normal routing.

Tool Contract

• May read `tool-capabilities-manifest.json`, selected skill metadata, Codex UI metadata when relevant, and local-only `project/mcp-profile.md`.
• May read or update `project/mcp-profile.md` when durable tool state is needed.
• May search official documentation only to verify installation sources or current client/tool docs.
• For active OpenAI questions, prefer a callable official OpenAI Developer Docs MCP provider, then the official web fallback; do not call either for generic frontend facts.
• Must prefer declared capabilities over server-name assumptions.
• Must not install MCP servers without explicit user approval.
• Must not change Codex, Claude, Cursor, VS Code, MCP, shell, package manager, or project configuration without explicit user approval.
• Must not run package installs.
• Must not access production systems or secrets.
• Route live design inspection to `design-screenshot-spec`; this skill maps

provider availability and does not acquire design content or edit canvases.

Core Tool Capabilities

Core capability vocabulary is defined in `common/tool-capability-model.md` and `tool-capabilities-manifest.json`.

Expected high-value capabilities:

project_files
command_execution
current_library_docs
web_platform_docs
rendered_visual_evidence
openai_platform_docs
client_platform_docs
repo_metadata
design_reference_files

Conditional design acquisition through `design-screenshot-spec`:

live_design_source

Prefer callable Figma MCP reads; fall back to browser/computer use with canvas pointer interaction and property-panel screenshots. Screenshot-only input needs no live provider. The capability excludes design writes and automatic setup.

Core does not mean always