🤖Free

Best MCP Servers for Claude Code: A Data-Backed Shortlist

Published: 2026-08-12Updated: 2026-08-12~9 min

MCP Radar Editorial · Facts verified: 2026-08-12 · Refresh due: 2026-09-12

Direct answer

There is no universal best Claude Code stack. Start with one server for a repeated gap—documentation, repository, browser testing, database, or project tracking—and add another only after measuring value and tool clutter. The ranking below is a dated shortlist: only two catalog samples have stored sandbox verification; most compatibility rows are transport-derived, not Claude Code tool-call tests.

Snapshot

2,636 records

Collected 2026-08-05

Claude rows

2,400

2 verified · 2,398 derived

Verified samples

2

Startup and tool list stored; not a full workflow review

Refresh

Monthly

Recompute after ownership, archive, or client changes

JobWhat to preferWhat to avoidVerification goal
Docs/contextNarrow source, current corpus, citationsHuge unfiltered context dumpsRetrieve one current API answer with provenance
RepositoryOfficial identity, scoped permissions, reversible readsBroad write/admin token by defaultRead one issue/PR before enabling writes
Browser/testingMaintained browser engine and explicit target scopeUnbounded profiles or saved sessionsOpen a local test page and return one observation
DatabaseRead-only credential and schema/query limitsProduction owner credentialsList schema and run a bounded read query
Project trackingPer-user auth and clear create/update scopesSilent external writesRead first; create only with confirmation
Choose by job and evidence. A listing is not an endorsement and a derived transport match is not a successful Claude Code invocation.

Start with the smallest useful stack

For most coding sessions, one or two high-value servers are enough. Claude Code already has repository and terminal capabilities, so an MCP server should close a repeated gap rather than duplicate a built-in path. Begin with documentation or one system of record, then add browser, database, or project tools when the workflow proves it needs them.

How the live shortlist is computed

The table is generated from the active catalog rather than hand-written. Candidates must match a developer workflow category, declare Claude Code compatibility, and clear the displayed adoption floor. Duplicate repositories are collapsed; eligible entries are ordered by TrustScore, then stars. TrustScore combines public maintenance, adoption, usability, health, and community signals. It is not a security score.

#ServerTrustScoreStarsClients
1io.github.D4Vinci/ScraplingScrapling MCP Server9373,830claude-code · derived
2io.github.microsoft/playwright-mcpPlaywright Tools for MCP8736,128claude-code · derived
3io.github.homeassistant-ai/ha-mcpComprehensive Model Context Protocol server for managing Home Assistant through 874,381claude-code · derived
4io.github.bytebase/dbhubDBHub873,365claude-code · derived
5io.github.ChromeDevTools/chrome-devtools-mcpChrome DevTools MCP8648,873claude-code · derived
6io.github.usekaneo/kaneoOfficial MCP server for Kaneo: manage tasks, projects, and labels from Claude an868,047claude-code · derived
7io.github.Ataraxy-Labs/semEntity-level code intelligence: semantic diff, impact analysis, blame, and conte863,302claude-code · derived
8io.github.atilaahmettaner/tradingview-mcpReal-time market data, screeners, technical analysis & backtesting for stocks, c853,937claude-code · derived
9io.github.zereight/gitlab-mcpGitLab MCP server for projects, merge requests, issues, pipelines, wiki, release851,886claude-code · derived
10io.github.github/github-mcp-serverGitHub8432,250claude-code · derived

10 of 2077 actively-maintained servers in these categories cleared the 500-star adoption floor, ordered by TrustScore. Regenerated daily from public GitHub, npm and registry signals — see our editorial policy for the full weighting. ✓ means we installed it in a sandbox and confirmed it starts.

Read the compatibility basis before the rank

The 2026-08-05 snapshot has 2,400 Claude Code compatibility rows. Only two are marked verified from stored sandbox startup and tool-list evidence; 2,398 are derived from a declared stdio or remote transport. Derived means 'configuration appears compatible,' not 'MCP Radar completed a Claude Code workflow.' Open each entity page and inspect the basis before installation.

Set up and smoke-test one pick

Add the server at local scope first, inspect it with `claude mcp list` and `claude mcp get`, open `/mcp`, and run one non-destructive task. Record the exact server identity, command/URL, auth scope, tools exposed, result, and date. Move a definition to project scope only after the team reviews the command and permission boundary.

Exclude servers that add more risk than leverage

Exclude a candidate when ownership or package identity is ambiguous, maintenance is stale, the requested credential is broader than the task, the same outcome is simpler through a trusted CLI, or the server exposes a large tool surface you will not use. A high rank cannot override a failed source, permission, or client test.

Methodology and limitations

  • Dataset frozen at 2026-08-05; client/docs checked 2026-08-12.
  • Ranking is computed from public signals after a category, lifecycle, adoption, and declared-client gate.
  • Two stored sandbox verifications prove startup and tool discovery only. The attempted current Claude Code health probe did not complete in the observation window and is not counted as a success.
  • Paid upstream APIs and vendor plans are not assumed to be free; verify pricing on the chosen server's primary source.

Frequently asked questions

How many MCP servers should I enable in Claude Code?

Start with one for a repeated gap and add only when the next server produces measurable value. More tools increase permission surface, selection ambiguity, and context overhead.

Does the ranking mean every server was tested in Claude Code?

No. The table requires a compatibility declaration, but most rows are transport-derived. The evidence basis is disclosed and only two catalog samples have stored sandbox startup/tool-list verification.

Should I share MCP config with my team?

Share a project-scoped definition only after reviewing the command, endpoint, source, and permissions. Keep live credentials out of `.mcp.json`.

Sources

  1. Claude Code — connect to tools via MCP · retrieved 2026-08-12
  2. Model Context Protocol — security best practices · retrieved 2026-08-12
  3. MCP Radar dataset snapshot (2026-08-05) · retrieved 2026-08-12

Continue reading