io.github.rzere/mcpexplorer

Official14 tools

by rzere

MCPExplorer

Search, vet & assemble MCP servers from your agent: verified tools, risk labels, and trust scores.

Search and evaluate MCP servers with verification, risk labels, and trust scores.

Topics

ai-agentsclaudellm-toolsmcpmcp-directorymcp-registrymcp-servermcp-serversmodel-context-protocol

Captured live from the server via tools/list.

search_servers

Search the MCPExplorer index of MCP servers by free text and filters. Omit `query` to browse the whole index ranked by trust score. Every filter either applies or is echoed back in `ignored_filters` with a reason — filters never silently do nothing.

Parameters (6)

querystring
Free-text query, e.g. 'send email' or 'postgres'. Omit to browse the full index ranked by trust.
official_statusstring
Filter by provenance
transportstring
Filter by transport. Many servers are still 'unknown' (transport not yet probed by the crawler).
risk_levelstring
Filter by tool risk. 'read-only' = has tools and none are write/destructive — the safe set to hand an autonomous agent.
capabilitystring
Filter to servers providing a capability slug, e.g. 'send-email'. An unknown slug is reported in ignored_filters, not silently dropped.
limitinteger
Max results (default 10)

get_server

Fetch the full record for one MCP server by slug: install methods, tools, trust, verification, and provenance.

Parameters (1)

slugstringrequired
Server slug, e.g. 'github'

get_capability

List the best MCP servers that provide a capability (e.g. 'send-email', 'web-search'), ranked by trust.

Parameters (1)

slugstringrequired
Capability slug, e.g. 'send-email'

compare_servers

Compare two MCP servers side by side: trust, transport, tools, install, and links.

Parameters (2)

astringrequired
First server slug
bstringrequired
Second server slug

generate_runtime_config

Generate the install + client config for an MCP server in a given runtime (claude-desktop, cursor, vscode, windsurf, cline, continue, goose, openai-agents, langgraph, crewai).

Parameters (2)

slugstringrequired
Server slug
runtimestringrequired
Runtime slug, e.g. 'claude-desktop'

check_trust

Return the explainable trust score, label, scoring reasons, and verification history for an MCP server.

Parameters (1)

slugstringrequired
Server slug

get_alternatives

Find verified MCP servers with the same capabilities as a given server, ranked by trust.

Parameters (1)

slugstringrequired
Server slug

list_loadouts

List curated loadouts — deliberately-assembled kits of MCP servers + governance + plays for a specific job (GTM, coding, research, support, infra). The agent-facing version of the /loadouts product. Use get_loadout for the full kit with live trust.

No parameters.

get_loadout

Fetch one curated loadout: its servers (each with role + LIVE trust/transport/tool-risk from the index), an aggregated governance posture for the whole kit, and the plays (see→decide→act sequences) it's designed to run.

Parameters (1)

slugstringrequired
Loadout slug, e.g. 'gtm', 'coding', 'research'

list_capabilities

List the capability slugs an agent can filter or look up by (e.g. 'send-email', 'query-database'), most-populated first. Use these with get_capability or search_servers' capability filter — the taxonomy isn't guessable.

Parameters (1)

limitinteger
Max capabilities (default 30)

get_ecosystem_stats

Live snapshot of the indexed MCP ecosystem: totals (servers, tools, handshake-verified), provenance and transport breakdowns, and the current top-trust / most-adopted / newest / biggest-trust-riser servers.

No parameters.

list_runtimes

List the runtimes generate_runtime_config supports (Claude Desktop, Cursor, VS Code, agent frameworks, …), with each one's config path. Enumerate these instead of guessing runtime slugs.

No parameters.

plan_toolset

Turn a plain-language goal into a recommended, governed toolset: the best-fit curated loadout (matched transparently by goal terms) plus an assembled set of the most relevant indexed servers, each with LIVE trust and tool-risk, and an aggregate governance posture for the kit. The decision layer over search/capability/trust/loadouts.

Parameters (3)

goalstringrequired
What the agent should do, in plain language, e.g. 'triage support tickets and file bugs'.
risk_tolerancestring
'read-only' restricts the assembled set to servers with no write/destructive tools — the safe set for an unsupervised agent. Default 'any'.
limitinteger
Max assembled servers (default 6).

report_broken_server

Report that an MCP server's install, docs, or endpoint is broken. Creates a submission for editorial review.

Parameters (2)

slugstringrequired
Server slug
detailstringrequired
What's broken (install fails, endpoint down, docs wrong, …)

README.md6 KBView on GitHub

MCPExplorer

The search engine and trust layer for MCP servers.

→ mcpexplorer.com

MCPExplorer indexes Model Context Protocol servers and verifies each one with a live handshake — so you can see the real tools a server exposes, each labeled by risk, plus a trust score, before you install it.

Every "awesome MCP" list is a static, unverified README. MCPExplorer actually runs the servers.

This repo is the public home for MCPExplorer's docs and its MCP server manifest. The product itself lives at mcpexplorer.com.

The problem

As the MCP ecosystem grew past thousands of servers, the lists stopped being useful. From a name and a one-line description you can't tell:

what tools a server actually exposes,
whether those tools are read-only or can delete your data,
whether it needs credentials,
or whether it's still maintained —

until you install it and hope.

What MCPExplorer does

Verifies, doesn't just list. It runs the real MCP tools/list handshake on every server — remote servers over the network, local/stdio servers in a hardened, ephemeral Docker sandbox — and extracts the true tool surface.
Labels tool risk. Every tool is tagged read / write / destructive, so you know what a server can do before you connect it.
Scores trust. A deterministic score from provenance, verification, adoption, and freshness — with every contributing factor shown on the page, not a black box.
Stays fresh. Every fact carries provenance and a "last checked" timestamp.

Currently ~1,000 servers indexed · 3,500+ tools extracted · 278 handshake-verified.

Start here

🔍 Explore the index — https://mcpexplorer.com/explore
⭐ Best MCP servers (ranked) — https://mcpexplorer.com/best-mcp-servers
📖 What is an MCP server? — https://mcpexplorer.com/what-is-an-mcp-server
📊 State of MCP report — https://mcpexplorer.com/reports

The MCPExplorer MCP server

MCPExplorer exposes its own MCP server, so you can search, vet, and assemble servers from inside your agent — without leaving your client. Hand plan_toolset a plain-language goal and get back a recommended, governed toolset.

Endpoint: https://mcpexplorer.com/mcp (Streamable HTTP, no auth)

Add it to Claude Code:

bash

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

Or add it to any MCP client:

json

{
  "mcpServers": {
    "mcpexplorer": {
      "type": "streamable-http",
      "url": "https://mcpexplorer.com/mcp"
    }
  }
}

Tools (14):

Tool	What it does
`search_servers`	Search the index by capability, transport, and risk
`get_server`	A server's verified tools, install command, and trust breakdown
`get_capability`	Servers that provide a given capability
`compare_servers`	Side-by-side comparison of two servers
`get_alternatives`	Servers similar to a given one
`check_trust`	Explainable trust score and reasoning for a server
`generate_runtime_config`	Ready-to-paste client config for a server
`list_capabilities`	The capability slugs you can search and look up by
`list_runtimes`	Runtimes `generate_runtime_config` supports
`get_ecosystem_stats`	Live totals + provenance/transport breakdowns
`list_loadouts`	Curated server kits for a job (GTM, coding, research, …)
`get_loadout`	One kit with live trust + aggregated tool-risk
`plan_toolset`	Plain-language goal → a recommended governed toolset
`report_broken_server`	Flag a stale or broken listing

How verification works

Every server flows through a crawler pipeline:

code

discover → fetch → parse → normalize → resolve → enrich →
classify → verify (live handshake) → score trust → decide indexability

The verify stage is the heart of it. Remote servers are probed over the network; local/stdio servers are executed in a locked-down container (--read-only, --cap-drop ALL, --security-opt no-new-privileges, no host mounts, no secrets, memory/CPU/PID limits, auto-removed) that runs only known launchers (npx/uvx/uv/pnpm/bunx). The handshake extracts the real tools/list, and each tool's risk is derived from its name and schema.

A page is only indexed once it has real MCP-server evidence and enough verified facts — so the index stays servers, not framework repos or link lists.

For agents & programmatic access

MCP server: https://mcpexplorer.com/mcp (see above)
llms.txt: https://mcpexplorer.com/llms.txt

Submit a server

Missing something? Submit it at https://mcpexplorer.com/submit — it enters the same verification pipeline as everything else.

FAQ

Is this affiliated with Anthropic? No. MCP is Anthropic's open protocol; MCPExplorer is an independent index built on top of it.

Is the trust score a black box? No. It's deterministic, and every +/− factor is shown on each server's page.

How do you run untrusted servers safely? Local/stdio servers only ever run inside the hardened, ephemeral container described above — never on the host, never with secrets.

<sub>MCPExplorer · mcpexplorer.com · a discovery, trust, and capability graph for MCP servers.</sub>

Install

claude_desktop_config.json

{
  "mcpServers": {
    "mcpexplorer": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcpexplorer.com/mcp"
      ]
    }
  }
}

Desktop config is stdio-only; this bridges via mcp-remote. Native remote: Settings > Connectors.