mcp-management

name	mcp-management
description	Configure and manage Model Context Protocol servers for external tool access
compatibility	>=1.4

MCP Management

Skill metadata: version "1.1"; license MIT; tags [mcp, servers, configuration, integration]; compatibility ">=1.4"; recommended tools [codebase, editFiles, fetch].

MCP (Model Context Protocol) is GA in VS Code v1.102+. Servers provide tools, resources, and prompts beyond built-in capabilities. Config in .vscode/mcp.json (workspace) or profile-level mcp.json (user).

When to use

• The user asks to configure, add, list, or check MCP servers
• You need to determine which MCP servers are available
• A task would benefit from an external tool not yet configured

Configuration locations

Location	Scope
.vscode/mcp.json	Workspace — shared via VCS
Profile-level mcp.json	User — all workspaces
settings.json "mcp" key	User/Workspace alternative
Dev container customizations.vscode.mcp	Per-container

Commands: MCP: Open Workspace Configuration, MCP: Open User Configuration

Server tiers

Tier	Servers	When	Config
Always-on	filesystem, git	Every project	Enabled by default
External	github, fetch	GitHub/web access needed	github uses VS Code OAuth; fetch needs no creds
Documentation	docs	Third-party libraries	Owned stdio server; queries DevDocs public API (devdocs.io)
Web search	duckduckgo	Public web search/fetch	Owned stdio server; SSRF-protected
Reasoning	sequential-thinking	Structured reasoning traces	Owned stdio server; in-memory state, no network

Available servers

Server	Tier	Transport	Purpose
@modelcontextprotocol/server-filesystem	Always-on	npx (stdio)	File operations within the workspace; supports OS-level sandboxing
mcp-git-server (owned)	Always-on	uvx (stdio, Python)	Git history, diffs, and branch operations; repo locked to workspace root
github/github-mcp-server	Credentials	HTTP remote (https://api.githubcopilot.com/mcp/readonly)	GitHub API — read-only by default; issues, PRs, Actions, Dependabot (toolset-restricted)
mcp-fetch-server (owned)	External	uvx (stdio, Python)	SSRF-protected HTTP fetch — blocks loopback, private (RFC-1918), and link-local/IMDS ranges
mcp-docs-server (owned)	Documentation	uvx (stdio, Python)	Library documentation from DevDocs (devdocs.io) — 794+ doc sets, all queries stay local
mcp-duckduckgo-server (owned)	Web search	uvx (stdio, Python)	DuckDuckGo search + SSRF-protected page fetch; max_results clamped to 20
mcp-sequential-thinking-server (owned)	Reasoning	uvx (stdio, Python)	Stateful sequential reasoning; in-memory only, no network calls

Removed (v3.2.0): @modelcontextprotocol/server-memory — replaced by VS Code's built-in memory tool (/memories/). Archived: @modelcontextprotocol/server-github (npm) — replaced by github/github-mcp-server HTTP remote. Archived: mcp-server-fetch (upstream) — replaced by owned mcp-fetch-server with SSRF protection. Archived: @modelcontextprotocol/server-sequential-thinking (npx) — replaced by owned mcp-sequential-thinking-server.

Stack-specific servers

Not included in base template. Add to .vscode/mcp.json by stack:

Stack	Server	Notes
Browser/UI testing	@playwright/mcp (Microsoft)	Removed from default template in v0.7.0. Use Path A (browser tools) or Path B (Playwright CLI) from the webapp-testing skill. Add manually if required.
PostgreSQL, SQLite, Redis	Search MCP Marketplace	Official reference servers archived; find maintained replacements
Docker	Search MCP Marketplace	Evaluate trust and permissions carefully
AWS	Search MCP Marketplace	Fine-grained IAM via ${env:}, never hardcode

Discover servers: code.visualstudio.com/mcp · registry.modelcontextprotocol.io · glama.ai · smithery.ai

Sequential Thinking (optional)

mcp-sequential-thinking-server (owned, uvx) — stateful sequential reasoning with thought branching, revision, and history. Included in the base template config; invoke via mcp_sequential-th_sequentialthinking. Use for complex planning tasks, architectural decisions, or any task requiring explicit thought backtracking.

MCP capabilities (GA since v1.102)

MCP servers can expose four capability types:

Capability	Description	Agent interaction
Tools	Functions the agent can invoke (e.g., query database, call API)	Agent calls tools directly
Resources	Data sources the agent can read (e.g., database schemas, config files)	Agent reads from # context menu
Prompts	Reusable prompt templates provided by the server	Available via / slash commands
MCP Apps	Interactive UI components (forms, visualisations, drag-and-drop)	Rendered inline in chat responses
Sampling	Server requests the agent to generate text on its behalf	Agent responds to server requests

Additional features: elicitations (server requests user input via the agent), MCP auth (OAuth/token flows for secure server connections).

Server discovery

• MCP gallery: In the Extensions view, search @mcp to browse and install servers directly (installs to user profile or workspace)
• MCP Marketplace: Browse and install servers from code.visualstudio.com/mcp
• Official registry: github.com/modelcontextprotocol/servers
• Community registries: mcp.so, glama.ai, smithery.ai
• Agent plugins: MCP servers can be bundled inside agent plugins (@agentPlugins in Extensions view)

Adding a new server

Before adding any MCP server:

1. Check if a built-in tool or existing MCP server already covers the need
2. Search the MCP Marketplace (code.visualstudio.com/mcp) and official registry
3. Check for npx vs uvx vs HTTP remote transport — prefer HTTP remote for officially hosted servers (no local process, OAuth-managed auth)
4. Add to .vscode/mcp.json (workspace) or profile mcp.json (user) with appropriate tier
5. For credentials-required stdio servers, use ${input:} or ${env:} variable syntax — never hardcode secrets; HTTP remote servers use VS Code's built-in OAuth where supported
6. For npx-based stdio servers on Linux/macOS, add sandboxEnabled: true with sandbox.filesystem.denyRead rules for credential directories (~/.ssh, ~/.gnupg, ~/.aws) as a defence-in-depth measure against prompt injection. Optionally add sandbox.network.allowedDomains to restrict outbound network access. Do not sandbox uvx-based servers: the VS Code sandbox proxy intercepts PyPI network access during the uvx launcher phase and triggers repeated domain-approval prompts that cannot be reliably suppressed via per-server allowedDomains. The M4 audit check enforces this by exempting servers with command == "uvx" automatically.
7. In consumer template repos, keep optional servers such as github, fetch, and docs present in .vscode/mcp.json but disabled by default until setup or update explicitly enables them
8. Agent files can declare least-privilege MCP access using the mcp-servers frontmatter field. Treat this as forward-compatible policy metadata: GitHub Copilot cloud agents document support for the field today, while local VS Code agent support may lag behind

Sandbox compatibility (Linux)

On immutable Linux distros (Fedora Atomic/Bazzite/Silverblue, NixOS) where /home symlinks to /var/home, bwrap rejects allowWrite paths. Detect: [[ "$(readlink -f /home)" != "/home" ]] && echo "immutable" || echo "standard". standard: use sandboxed config. immutable: omit sandboxEnabled, sandbox, and top-level sandbox block.

Auto-start

Set "chat.mcp.autostart": "newAndOutdated" in .vscode/settings.json — servers start on chat message. Trust dialog shown on first auto-start.

CLI and external agent access

VS Code 1.113+: .vscode/mcp.json servers are bridged to Copilot CLI and Claude agents automatically.

Settings Sync

Enable Settings Sync: Configure → MCP Servers to sync configs across devices.

Subagent MCP use

Subagents inherit all configured MCP servers. To add a new server, the subagent flags the proposal to the parent agent before modifying .vscode/mcp.json.