---
configs:
- config_name: deduplicated
  data_files:
  - split: deduplicated
    path: data/deduplicated-*.parquet
- config_name: default
  data_files:
  - split: deduplicated
    path: data/deduplicated-*
- config_name: raw
  data_files:
  - split: raw
    path: data/raw-*.parquet
dataset_info:
  features:
  - name: name
    dtype: string
  - name: versions
    dtype: string
  - name: capabilities
    dtype: string
  - name: last_seen
    dtype: string
  splits:
  - name: deduplicated
    num_bytes: 2757
    num_examples: 32
  download_size: 3610
  dataset_size: 2757
---

# MCP Clients Dataset

This data contains clients logged by the MCP Server at `huggingface.co/mcp`.

## Dataset Contents

The dataset contains unique MCP client configurations with the following fields:

### Raw Split (`raw`)
- **name**: MCP client name (e.g., "Cursor", "Anthropic/ClaudeAI", "chat-ui-mcp")
- **version**: Client version
- **capabilities**: Client capabilities (JSON string)
- **last_seen**: Most recent timestamp when this client was seen

### Deduplicated Split (`deduplicated`)
- **name**: MCP client name (normalized, mcp-remote suffix stripped)
- **versions**: All versions seen for this client (comma-separated)
- **capabilities**: Client capabilities (JSON string)
- **last_seen**: Most recent timestamp when this client was seen

## Notes about the Data

- Some Clients only send capabilities that have been configured (for example, `fast-agent` only sends the `roots` capability if they have been configured)
- There are a number of capabilities out of specification that are helpful to track (e.g. MCP-UI, Skybridge, Apps SDK). We encourage the use of `experimental` to advertise these capabilities for tracking.
- The data is extracted from the `evalstate/hf-mcp-logs` dataset, filtering for `initialize` method calls and tracking the most recent `last_seen` timestamp for each unique client configuration.
- The `deduplicated` split groups clients by (name, capabilities), concatenating all versions together.

## Usage

```python
from datasets import load_dataset

# Load raw clients
raw_ds = load_dataset('evalstate/mcp-clients', 'raw')
for client in raw_ds['raw']:
    print(f"{client['name']} v{client['version']}")

# Load deduplicated clients
dedup_ds = load_dataset('evalstate/mcp-clients', 'deduplicated')
for client in dedup_ds['deduplicated']:
    print(f"{client['name']}: {client['versions']}")
```

## Updating the Dataset

The dataset is kept up-to-date via Hugging Face Jobs:

1. **Extract step** (`raw` split):
   ```bash
   uv run https://huggingface.co/datasets/evalstate/mcp-clients/resolve/main/extract_mcp_clients.py --push-to-hub --split raw
   ```

2. **Deduplicate step** (`deduplicated` split):
   ```bash
   uv run https://huggingface.co/datasets/evalstate/mcp-clients/resolve/main/deduplicate_clients.py --push-to-hub
   ```

## License

MIT