Building Your Tool Stack
- Design a personal tool stack using user scope for cross-project tools and project scope for team-shared tools
- Chain multiple connected tools in a single prompt to execute a multi-system workflow without switching applications
- Maintain a healthy stack by removing unused servers and managing the context budget as your tool count grows
The proliferation problem: twelve servers you forget to maintain
The natural trajectory of MCP adoption is to add servers eagerly. You connect GitHub on Monday, then Sentry on Tuesday, then Notion and Firecrawl and Playwright by the end of the week. Two months later you have twelve servers configured, four of them with expired OAuth tokens, two with API keys that rotated, and three you have not used in weeks. Every session starts by waiting for reconnection attempts that fail silently.
The goal is not the largest possible tool stack. It is the most effective minimal stack: the tools you actually use regularly, configured correctly, with credentials that stay valid. A focused set of five well-maintained servers produces better results than twelve half-working ones.
User scope versus project scope
The scope you choose for each server determines who has access to it and where the configuration lives.
--scope user)
Stored in ~/.claude.json. Available in every project on your machine. Private to you. Best for: personal credentials (your GitHub PAT, your Sentry account, your personal Notion workspace).
.mcp.json)
Stored in the project root. Version-controlled. Available to all teammates who clone the repo. Best for: shared tools (a team read-only database, a shared Sentry organization, a team Figma file).
The distinction matters most for credentials. Your GitHub PAT is personal — it accesses your account. A read-only database user shared across the team is project-level. Personal credentials belong in user scope; shared tools belong in project scope with credentials managed through environment variables.
Designing your personal stack
Start by listing the three to five tools you open most often during a coding session. For most developers it is some combination of:
- Code collaboration: GitHub (issues, PRs, code search)
- Issue tracking: Linear or Jira (ticket-to-code workflow)
- Error monitoring: Sentry (production debugging, release tracking)
- Knowledge base: Notion (specs, ADRs, onboarding docs)
- Database: Your application's read-only replica (data questions)
Add these at user scope so they are available in every project without reconfiguring them:
claude mcp add --scope user --transport http github https://api.githubcopilot.com/mcp/ --header "Authorization: Bearer YOUR_PAT"
claude mcp add --scope user --transport http linear https://mcp.linear.app/mcp
claude mcp add --scope user --transport http sentry https://mcp.sentry.dev/mcp
The multi-tool workflow: chaining connected servers
The real power of a connected tool stack is the ability to chain multiple systems in a single prompt. The official Claude Code documentation describes the full version of this workflow:
Add the feature described in Linear issue ENG-4521. Check Sentry and Statsig to check the usage of that feature. Find emails of 10 random users who used the feature based on our PostgreSQL database. Create Gmail drafts inviting them to a feedback session.
That prompt crosses Linear, Sentry, PostgreSQL, and Gmail. With all four connected, Claude executes the full chain: reads the ticket, checks error rates and usage data, queries the database, and drafts the invitations — without switching applications once. Without the connected tools, executing that workflow manually takes 20–30 minutes of tab switching and copy-pasting.
Other tools worth knowing from the Directory
Beyond the tools covered in this track, the Anthropic Directory (claude.ai/directory) lists many more verified integrations. A few worth noting for specific use cases:
- Stripe (
https://mcp.stripe.com) — query subscription status, review recent transactions, check payment failures. Useful for SaaS developers who need to correlate user behavior with billing state. - Airtable (
npx airtable-mcp-serverwithAIRTABLE_API_KEY) — read and write Airtable bases. Useful when your team uses Airtable for structured data that does not live in a traditional database. - HubSpot (
https://mcp.hubspot.com/anthropic) — read CRM data, contact records, and deal status. Useful for developers building integrations with marketing and sales systems.
Maintaining your stack
A healthy stack requires occasional maintenance:
Remove what you do not use. Every connected server loads tool schemas into Claude's context window. A server you use once a month is not worth the context overhead in every session. Remove it with claude mcp remove SERVER_NAME and re-add it when you need it.
Rotate credentials on a schedule. GitHub PATs, Atlassian API tokens, and Firecrawl API keys have expiration dates. Set a calendar reminder to rotate them before they expire — a failed server is less obvious than an error message, and Claude may proceed without the tool rather than telling you it is unavailable.
Use claude mcp list to audit. Run it periodically to see which servers are connected, which need re-authentication, and which have been failing silently. A server showing ✗ Failed to connect is not helping Claude help you — fix it or remove it.
- User scope (<code>--scope user</code>) is for personal credentials available across all your projects; project scope (<code>.mcp.json</code>) is for team-shared tool configurations that belong in version control.
- The most effective stacks are minimal and maintained: three to five tools you actually use daily, with valid credentials, outperform twelve rarely-used servers with stale tokens.
- Multi-tool prompts that chain Linear, Sentry, a database, and Gmail in a single session represent the full value of connected tools — the copy-paste loop across four applications becomes one prompt.
- Run <code>claude mcp list</code> periodically to audit which servers are connected, which need re-authentication, and which are failing silently.
- Remove servers you rarely use — every connected server loads tool schemas into Claude's context window, and a focused stack leaves more budget for the actual work.