Security review

Is magic-mcp MCP safe to give an agent?

read onlyneeds cautionhealthyhealthy

A factual risk summary built from magic-mcp’s real tool surface, execution model, and verification history — not a vibe. Trust score 70/100.

01What it can do

Only read-style tools observed — no write/destructive tools.

4 tools observed

02Execution model

Runs locally over stdio — the server process executes on your machine with your user's privileges. Vet the source and package before granting access.

Installs via npx (`npx -y @21st-dev/magic@latest`) — it pulls and executes third-party code; pin versions and review the source.

03Permissions & auth

Manage GitHub · writeSend chat messages · write

The write/destructive access this server can exercise, inferred from its verified tools.

04Verification

handshakenot confirmed

runshandshake:skipped · metadata:passed · handshake:skipped · metadata:passed · handshake:skipped · metadata:passed · handshake:skipped · metadata:passed · handshake:skipped · metadata:passed

last_checked2026-08-16 21:33Z

sourcesGitHub repo search [p4] · Awesome MCP Servers (wong2) [p6]

Reduce the risk

Worried about handing an agent raw access? See governed agents in action — Apex gives your AI paced, capped, fully-logged hands with approval queues before anything runs.

Explore Apex →

See also: full server page · setup · alternatives