automatelab/mcp-servers-tool-catalog

MCP Servers Tool Catalog

Machine-introspected catalog of public Model Context Protocol (MCP) servers and the tools each one exposes. Each tool row contains the exact name, description, and JSON Schema input_schema captured by running the server and calling its tools/list JSON-RPC endpoint -- not metadata scraping.

Snapshot packaged: 2026-10-12 Refresh coverage: Only six AL-owned servers were re-probed. Other records retain their previous observations. Observation manifest date: 2026-10-12T14:07:09.314715+00:00 Servers attempted: 926 Servers successfully introspected: 363 Tools captured: 9971

Why this dataset exists

The MCP ecosystem grew from ~1,200 servers in Q1 2025 to ~9,400 servers by April 2026, with 97M monthly SDK downloads. Every AI agent shipping MCP support (Claude Code, Cursor, Cline, Zed) faces the same problem: choosing servers blind. Existing datasets are insufficient:

Dataset	Servers	Tool defs?	Refresh	Downloads/mo
ScaleAI/MCP-Atlas	36	yes (benchmark subset)	static	2,588
DeepNLP/mcp-servers	296	metadata only	stale since 2025-03	~20
automatelab/mcp-servers-tool-catalog	926	yes (full schemas)	monthly	--

Three differentiators:

1. Actual tool introspection -- we run each server and capture its tools/list response, not its README.
2. Explicit coverage of attempted servers, with failures retained alongside successful probes.
3. Monthly refresh with diff-based changelog (vs. static snapshots).

Schema

tools split

field	type	description
server_name	string	short slug for the server
package	string	exact npm package name
tool_name	string	name field from tools/list
tool_description	string	description field from tools/list
input_schema	string (JSON)	inputSchema field, JSON-encoded for parquet
schema_validation_status	string	Metaschema check result; invalid recorded schemas are preserved, not repaired or dropped
schema_dialect	string	Declared JSON Schema dialect, or an explicit unspecified value

servers split

field	type	description
name	string	short slug
package	string	npm package
status	string	ok, or one of the failure categories below
elapsed_s	float	wall-clock seconds for introspection
tool_count	int	tools captured (0 if not ok)
server_info	string (JSON)	server's response to initialize
stderr_snippet	string	last meaningful stderr line if introspection failed
raw_status	string	underlying state when classification couldn't pinpoint cause

Failure categories

Each non-ok row has a status from the following set, derived by matching the server's stderr against known patterns:

• npm_404_not_found -- package does not exist on the npm registry
• npm_network_error -- transient registry/network issue at install
• needs_<vendor>_token / needs_<vendor>_key -- server hard-requires real credentials at startup (e.g. needs_stripe_key, needs_github_token)
• needs_database_url / needs_aws_creds / needs_azure_creds / needs_google_creds -- backing-service credentials required
• needs_config_file -- server expects a config file path
• needs_cli_args -- server requires positional CLI arguments (e.g. a root path)
• needs_python -- runtime depends on Python but python is not on PATH
• broken_install -- npm install completed but the server fails to load (missing module, etc.)
• init_timeout -- spawned, but did not respond to initialize within timeout
• tools_list_timeout -- responded to initialize but not to tools/list

Methodology

For each candidate package:

1. Verify existence with npm view <pkg> name.
2. Spawn npx -y <pkg> over stdio with a JSON-RPC client.
3. Send initialize (protocol version 2024-11-05) and notifications/initialized.
4. Send tools/list, capture the response.
5. Classify stderr against ~17 known failure patterns if any step fails.
6. Write one row per server (always) and one row per tool (only on success).

Auth-gated servers were given dummy environment variables (STRIPE_SECRET_KEY=dummy, etc.). Most servers that list tools without hitting their backing service worked fine with dummies; those that validate credentials at initialize time get needs_<vendor>_* status.

The pipeline source is open: see introspect.py and validate_npm.py in the dataset repo.

License

Dataset additions: CC-BY-4.0. Each tool row attributes its source package to the upstream MCP server repository under its own license (see servers split).

Refresh cadence

Monthly. Each version tag includes a diff vs. the previous month (added servers, removed servers, new tools per server, removed tools).

How to load

from datasets import load_dataset
tools = load_dataset("automatelab/mcp-servers-tool-catalog", "tools", split="train")
servers = load_dataset("automatelab/mcp-servers-tool-catalog", "servers", split="train")

Acknowledgements

Built by automatelab. Companion site: automatelab.tech/products/datasets/mcp-tool-catalog/. Pipeline source: github.com/AutomateLab-tech/mcp-tool-catalog.

Snapshot provenance and limitations

926 attempted servers, 363 successful probes and 9,971 recorded tools. This refresh re-probed six owned servers only; other records retain their previous observation dates. Failed probes are explicit coverage gaps.