README.md

---
license: cc-by-4.0
language:
  - en
tags:
  - mcp
  - model-context-protocol
  - ai-agents
  - ai-safety
  - tool-classification
  - security
  - policy
size_categories:
  - 10K<n<100K
pretty_name: PolicyLayer MCP Server Catalogue
configs:
  - config_name: default
    data_files:
      - split: servers
        path: servers.jsonl
      - split: tools
        path: tools.jsonl
---

# PolicyLayer MCP Server Catalogue

A risk-classified catalogue of every Model Context Protocol (MCP) server reachable through the public registries. Each tool exposed by each server is classified into one of six risk categories using a verb-based classifier with input-schema heuristics.

**1,787 servers · 25,329 tools · classified May 2026 · refreshed monthly.**

The catalogue is the underlying dataset for PolicyLayer's research report,
[*The State of MCP Security — May 2026*](https://policylayer.com/research/state-of-mcp-2026).

---

## Loading the dataset

```python
from datasets import load_dataset

ds = load_dataset("PolicyLayer/mcp-server-catalogue")
servers = ds["servers"]
tools   = ds["tools"]

print(f"{len(servers)} servers, {len(tools)} tools")

# Find every destructive tool exposed by an identity-provider MCP server
ip_slugs = [s["slug"] for s in servers if "identity" in (s["description"] or "").lower()]
destructive = tools.filter(
    lambda t: t["category"] == "Destructive" and t["server_slug"] in ip_slugs
)
```

---

## Schema

### `servers` split

One row per Model Context Protocol server with at least one classified tool.

| Field | Type | Description |
|---|---|---|
| `slug` | string | Stable identifier. Use this to join with `tools.server_slug`. |
| `name` | string | Human-readable server name as published by its author. |
| `packages` | array of strings | npm or other package identifiers exposing this server. |
| `repo_url` | string \| null | Source repository URL where available. |
| `description` | string \| null | Server-author-supplied description. |
| `tool_count` | int | Number of tools the server exposes. |
| `categories` | array of strings | Distinct risk categories present in this server's tools. Subset of {`Read`, `Write`, `Execute`, `Destructive`, `Financial`, `Other`}. |
| `source` | string | Discovery channel: `crawler` (npm auto-discovery), `smithery` (Smithery registry), `seed` (manually-added official), `user_scan` (community contribution). |
| `confidence` | string | Classifier confidence for this server's labels: `high`, `verified`, `medium`, `low`. |
| `last_scanned_at` | string (ISO 8601) | Most recent scan timestamp. |
| `npm_weekly_downloads` | int \| null | Weekly downloads of the server's npm package as of the snapshot date. Null if the server is not on npm or the registry didn't return data. |
| `npm_latest_version` | string \| null | Latest published version on npm. |
| `npm_last_published` | string (ISO 8601) \| null | Timestamp of the latest npm publish. |
| `github_stars` | int \| null | GitHub stargazer count for the source repository. |
| `github_forks` | int \| null | GitHub fork count. |
| `github_open_issues` | int \| null | Currently-open GitHub issues + pull requests. |
| `github_last_commit` | string (ISO 8601) \| null | `pushed_at` timestamp on the GitHub repository — proxy for liveness. |
| `github_license` | string \| null | SPDX licence identifier reported by GitHub (e.g., `MIT`, `Apache-2.0`). |
| `github_archived` | bool \| null | Whether the GitHub repository is archived (read-only). |

### `tools` split

One row per tool exposed by a server in the `servers` split.

| Field | Type | Description |
|---|---|---|
| `server_slug` | string | Foreign key to `servers.slug`. |
| `name` | string | Tool name as exposed via `tools/list`. |
| `description` | string \| null | Tool-author-supplied description. |
| `category` | string | Risk category. One of `Read`, `Write`, `Execute`, `Destructive`, `Financial`, `Other`. |
| `severity` | string | Discrete severity bucket: `Low`, `Medium`, `High`, `Critical`. |
| `confidence` | string | Classifier confidence: `high`, `verified`, `medium`, `low`. |
| `risk_weight` | float | Risk weight in [0.0, 1.0]. 0.0 = read-only, 1.0 = destructive financial. |
| `risk_type` | string \| null | Specific risk type where determinable (e.g., `delete`, `drop`, `payment`). |
| `input_schema` | string \| null | JSON-encoded tool input schema. May be `null` if the source provided no schema. |

---

## Risk categories

| Category | Definition | Risk weight typical range |
|---|---|---|
| `Read` | Retrieves data without modification. | 0.05–0.15 |
| `Write` | Creates or modifies data and resources. | 0.20–0.45 |
| `Execute` | Runs code, scripts, or container commands. | 0.40–0.70 |
| `Destructive` | Permanently deletes or revokes resources. | 0.55–0.85 |
| `Financial` | Moves real money: charges, payments, refunds, transfers. | 0.65–1.00 |
| `Other` | Does not fit the above categories. | varies |

Classification is verb-based with input-schema heuristics. Methodology details and known failure modes are documented in the
[research report](https://policylayer.com/research/state-of-mcp-2026#methodology).

---

## Coverage and limitations

- **Public registries only.** The dataset covers servers reachable through the official Model Context Protocol registry, npm, Smithery, and Glama. Private and self-hosted servers are not represented.
- **Lower bound.** Some registry-listed servers are unreachable through the scan pipeline (broken installs, dependency failures, or auth-walled launchers) and are excluded. The dataset is therefore a lower bound on the real ecosystem.
- **Static snapshot.** This release is the May 2026 edition. New monthly versions are released on the 1st of each month.
- **Classifier confidence varies.** 72.3% of tool classifications are high-confidence and 15.4% are verified. The remaining 12.3% are medium or low confidence and should be treated as advisory rather than authoritative for downstream risk decisions.
- **Usage metrics are point-in-time.** `npm_weekly_downloads` reflects the week immediately before the snapshot. `github_*` fields reflect the GitHub repository state at fetch time. Metrics are null where the server is not on npm, has no GitHub repo, or the upstream API rate-limited the fetch.

---

## Citation

If you use this dataset in research or commentary, please cite the research report:

```
PolicyLayer. (2026). The State of MCP Security — May 2026.
https://policylayer.com/research/state-of-mcp-2026
```

For the dataset itself:

```
PolicyLayer. (2026). PolicyLayer MCP Server Catalogue [Data set].
Hugging Face. https://huggingface.co/datasets/PolicyLayer/mcp-server-catalogue
```

---

## Updates

The catalogue is regenerated monthly from PolicyLayer's continuously-updated scan pipeline. Each monthly release is tagged on Hugging Face. Subscribe to the dataset to be notified of new versions.

| Edition | Tag | Servers | Tools |
|---|---|---:|---:|
| May 2026 | `2026-05` | 1,787 | 25,329 |

---

## Related

- **Research report:** [The State of MCP Security — May 2026](https://policylayer.com/research/state-of-mcp-2026)
- **Public tool catalogue:** [policylayer.com/tools](https://policylayer.com/tools) — browse classifications by server.
- **MCP attack database:** [policylayer.com/attacks](https://policylayer.com/attacks) — documented MCP incidents and attack patterns.
- **PolicyLayer:** [policylayer.com](https://policylayer.com) — the MCP control plane (gateway, policy engine, audit log) for production agent fleets.

Questions, methodology details, or data requests: research@policylayer.com.

---

## Licence

This dataset is released under [Creative Commons Attribution 4.0 International (CC-BY-4.0)](https://creativecommons.org/licenses/by/4.0/).

You are free to use, share, adapt, and redistribute the data, including for commercial purposes, with attribution to PolicyLayer. The classifier output is the work product of PolicyLayer's research team; tool descriptions and server metadata are the work of their respective authors and remain under whatever licence the source projects supply.