Best MCP servers for Claude Code

Claude Code already reads your files, runs your tests and uses git. MCP servers add the context it can't get from the repository: the issue it is working on, the error that happened in production, the current docs for a library and the deploy that just failed. The servers below are the remote, vendor-run ones that help most with coding work.

Each is an HTTPS endpoint published by the vendor or a well-known publisher, and each was live-probed on 2026-10-29. All of them connect with one claude mcp add command. For the broader list across product, support and marketing tools, see the best remote MCP servers.

Adding a remote server to Claude Code

The syntax from Claude Code's MCP docs is:

claude mcp add --transport http <name> <url>

The default scope is local, which loads the server only in the current project. Add --scope user to make it available everywhere, or --scope project to write it to a shared .mcp.json. For OAuth servers, start Claude Code, run /mcp, pick the server and finish sign-in in the browser. For token-based servers, pass --header "Authorization: Bearer <token>" after the URL.

The picks

1. GitHub

The most common reason to add an MCP server to a coding agent. Claude Code can already push with git, but the GitHub server lets it read issues and review comments, search code across repositories and open or update pull requests.

claude mcp add --transport http github https://api.githubcopilot.com/mcp/ \
  --header "Authorization: Bearer YOUR_GITHUB_PAT"

• Try: "Read the review comments on PR #214, fix the ones that are about code, and reply to each with what you changed."
• Permission tip: use a fine-grained token limited to the repositories you work in. Let Claude Code read freely, but don't pre-approve merge, delete or repository settings tools.

2. Linear

The Linear server finds, creates and updates issues, projects and comments. Give Claude Code an issue ID and it can read the description and discussion before touching code.

claude mcp add --transport http linear https://mcp.linear.app/mcp

• Try: "Pick up ENG-311. Read the issue and comments, implement it, then comment on the issue with a summary and the branch name."
• Permission tip: Linear offers https://mcp.linear.app/mcp/readonly if you only want the agent to read issues. On the full endpoint, approve status changes yourself.

3. Sentry

The Sentry server searches errors and pulls event context, so Claude Code sees the stack trace, tags and breadcrumbs for the bug it is fixing. Scope the URL to your organization and project to keep the tool list short.

claude mcp add --transport http sentry https://mcp.sentry.dev/mcp/my-org/my-project

• Try: "Find the most frequent unresolved error in the last 24 hours, locate the code path in this repo and propose a fix with a regression test."
• Permission tip: reads are safe to pre-approve. Keep tools that resolve or reassign issues behind a prompt.

4. Context7

Context7, from Upstash, returns current documentation and examples for a library. Its two tools, resolve-library-id and query-docs, are read-only. No sign-in is needed; Upstash offers a free API key for higher rate limits.

claude mcp add --transport http context7 https://mcp.context7.com/mcp

• Try: "Upgrade our TanStack Router usage to the current API. use context7"
• Permission tip: pre-approve both tools. Add a line to CLAUDE.md such as "Use Context7 when you need library or API documentation" so Claude Code reaches for it without being asked.

5. DeepWiki

DeepWiki, from Cognition, answers questions about public GitHub repositories using generated wikis. It is free and needs no sign-in. Tools are read_wiki_structure, read_wiki_contents and ask_question.

claude mcp add --transport http deepwiki https://mcp.deepwiki.com/mcp

• Try: "Using DeepWiki, explain how vercel/next.js decides between static and dynamic rendering, then check whether our layout opts out by accident."
• Permission tip: read-only and public data, so pre-approve it. It covers public repositories only.

6. Supabase

The Supabase server inspects schemas, runs SQL and works with migrations, branches and other project resources. The URL takes project_ref, read_only and features parameters, and using them is the simplest way to keep the agent away from the wrong project.

claude mcp add --transport http supabase \
  'https://mcp.supabase.com/mcp?project_ref=YOUR_PROJECT_REF&read_only=true'

• Try: "Look at the schema for subscriptions and write a migration that adds a canceled_at column with a backfill. Don't apply it."
• Permission tip: Supabase recommends keeping manual approval of tool calls on. Drop read_only=true only on a development project, and never pre-approve SQL execution or migration tools.

7. Neon

The Neon server creates projects and branches, runs SQL and helps plan migrations on Neon Postgres. Branching is the useful part: the agent can test a migration on a copy of the database.

claude mcp add --transport http neon 'https://mcp.neon.tech/mcp?readonly=true'

• Try: "Create a branch from main, apply this migration there and run the test queries against it."
• Permission tip: Neon says to use its MCP server for development and testing, not production. Remove readonly=true only for that, and add projectId=<id> to limit it to one project.

8. Vercel

The Vercel server searches Vercel docs, finds projects and deployments, reads build logs and queries Web Analytics. Vercel lists Claude Code as a supported client.

claude mcp add --transport http vercel https://mcp.vercel.com

• Try: "The last preview deployment failed. Read the build log, fix the cause in this branch and push."
• Permission tip: connecting gives the agent your Vercel user's access. The server can deploy, so leave deploy tools behind a prompt.

9. Railway

The Railway server inspects projects, services and deployments and can redeploy. Railway supports OAuth directly at its endpoint, where you choose which workspaces and projects the client can reach.

claude mcp add --transport http railway https://mcp.railway.com

• Try: "Why is the worker service in staging crash-looping? Check the latest deployment and tell me what changed."
• Permission tip: Railway flags destructive tools and calls out redeploy and accept-deploy as ones to review. Don't pre-approve them.

10. Datadog

The Datadog server searches logs, metrics and traces and reviews monitors, dashboards and incidents, using your own Datadog permissions. The endpoint below is for the US1 site; other Datadog sites use their own hostname.

claude mcp add --transport http datadog \
  https://mcp.datadoghq.com/api/unstable/mcp-server/mcp

• Try: "Find the p95 latency spike on checkout-api this morning, pull the slow traces and point me to the code they spend time in."
• Permission tip: Datadog lets you limit tools at connection time with toolsets and checks write permissions on every call. Pre-approve log and trace search; review anything that changes monitors.

11. Postman

The Postman server manages collections and environments and inspects APIs and specifications. It helps when the API you are changing is documented in Postman rather than in the repo.

claude mcp add --transport http postman https://mcp.postman.com/mcp

• Try: "Update the Orders collection to match the new POST /orders request body in this branch."
• Permission tip: this endpoint exposes Postman's full tool set, which Postman says is over 100 tools. Postman also documents smaller tool sets; use one if you only need a few operations, so the tool list doesn't crowd the context.

Connecting them all through Toolstile

Eleven claude mcp add commands work, but they live in one machine's config, have to be repeated in Codex, Cursor or a second laptop, and give Claude Code every tool each server offers. If you use more than a couple of these, a gateway keeps that in one place.

With Toolstile, you connect GitHub, Linear, Sentry and the rest once in a workspace. Then Claude Code gets one entry:

claude mcp add --transport http toolstile 'https://toolstile.com/mcp'

Start Claude Code, run /mcp, select Toolstile and sign in through the browser to choose a workspace. Full steps for each client are in connect your agent, and getting started covers adding the first connection. The integrations directory lists the server URL for every server above.

The new Claude Code authorization starts with no access. In Toolstile you choose, per tool, whether it is Allowed, needs approval or is Blocked. Claude Code on your laptop can comment on pull requests and read Sentry, while Codex on a CI runner only reads checks, even though both use the same GitHub connection. Blocked tools don't appear in the agent's tool list, which also keeps the context smaller.

Claude Code's own permission settings still apply on top. Claude Code permissions for MCP tools covers how the two layers fit together, and permissions explains the Toolstile side.

FAQ

Do I need all of these?

No. Start with your issue tracker, your code host and Context7. Add Sentry or Datadog when you debug production issues, and a database server only when you work on schemas. Every server's tool definitions are sent to the model, so unused servers cost context.

Which ones work without an account?

Context7 and DeepWiki need no sign-in. The others use OAuth through /mcp, except GitHub, where a personal access token in a header is the documented Claude Code setup.

Can I share these servers with my team in Claude Code?

--scope project writes the server to .mcp.json in the repository, so teammates get the same entry and sign in with their own accounts. Don't commit tokens in headers. For shared access rules across apps and machines, use a gateway as described above.