# Nerf Watch

> Nerf Watch is a free MCP server for ai models and llm tooling. Daily scores for Claude, GPT and other models on one private, code-graded test suite. It runs as a hosted remote server, so there is nothing to install locally.

- **Kind:** MCP server
- **Pricing:** Free
- **Platforms:** API, Web
- **Website:** https://tokenusage.dev/nerfwatch
- **Repository:** https://github.com/stdcmms-lang/tokenusage-codex-mcp/tree/HEAD/nerfwatch
- **MCP Registry name:** dev.tokenusage/nerfwatch
- **Categories:** AI models and LLM tooling
- **Listing updated:** 2026-11-14

## About

Nerf Watch tracks whether coding models get worse over time. Each day the same private tests run against Claude models through Claude Code, OpenAI models through Codex, and several others through the OpenCode CLI, with code doing the grading. The questions stay private to limit training on them, and a marker appears on a model's chart when a score jumps more than 10 points between runs.

The MCP server exposes the same public data read-only. One tool returns latest scores with previous runs, changes, 7-day baselines and trends, including model ids. The other takes a model id and returns its retained history, oldest first, with the suite version for each point. Missing days are missing rather than zero, and comparisons are made within one suite version. Test questions and grader details are not returned.

The README cautions that a lower score is a lower score on this suite, not proof that a provider changed a model, since runs are noisy. Both tools share a 300-second cache, and shown data needs a `Data: tokenusage.dev` credit. The service is unofficial.

## Features

- **Latest scores:** Returns each model's current score, change from the last run and 7-day baseline.
- **Model history:** Returns one model's retained scores, oldest first, with suite version.
- **Trend labels:** Scores within 2 points of the baseline read as steady.
- **Private suite:** Test questions stay hidden to reduce training on them.
- **No key needed:** Hosted, read-only, no account.
- **Speed probe:** The site also shows daily response times by model.

## Good for

- Ask whether a model's quality dipped this week
- Compare Claude and GPT models on the same suite
- Chart a model's score history in a dashboard
- Check if a score change is within the usual noise

## Install

### Claude Code (Terminal)

```sh
claude mcp add --transport http nerfwatch https://tokenusage.dev/mcp/nerfwatch
```

### Claude Desktop (Settings → Connectors → Add custom connector)

```sh
https://tokenusage.dev/mcp/nerfwatch
```

### Cursor (~/.cursor/mcp.json)

```json
{
  "mcpServers": {
    "nerfwatch": {
      "url": "https://tokenusage.dev/mcp/nerfwatch"
    }
  }
}
```

### VS Code (.vscode/mcp.json)

```json
{
  "servers": {
    "nerfwatch": {
      "type": "http",
      "url": "https://tokenusage.dev/mcp/nerfwatch"
    }
  }
}
```

### Windsurf (~/.codeium/windsurf/mcp_config.json)

```json
{
  "mcpServers": {
    "nerfwatch": {
      "serverUrl": "https://tokenusage.dev/mcp/nerfwatch"
    }
  }
}
```

### Other clients (your client's MCP settings (most use this shape))

```json
{
  "mcpServers": {
    "nerfwatch": {
      "type": "http",
      "url": "https://tokenusage.dev/mcp/nerfwatch"
    }
  }
}
```

## Tools

- `get_nerfwatch_scores`: Returns the latest public scores, previous runs, changes, baselines and trends with model ids.
- `get_nerfwatch_history`: Returns one model's retained score history, oldest first, with suite versions.

## FAQ

### Does a lower score mean a model was changed?

No. The README says it is only a lower score on this suite, as runs are noisy.

### What do I need to use it?

No key or account.

### How often is it updated?

Tests run once a day, and responses are cached for 300 seconds.

### Can I show the data publicly?

Yes, with a visible Data: tokenusage.dev credit linking to the Nerf Watch page.

### Is Nerf Watch free?

Yes. Nerf Watch is free to use.

### How do I add Nerf Watch to Claude Code?

Run this command in your terminal: claude mcp add --transport http nerfwatch https://tokenusage.dev/mcp/nerfwatch

### What are the best alternatives to Nerf Watch?

The closest alternatives to Nerf Watch are Claude Code Ultimate Guide, signafAIde, Toolkien, Lugh and HypePaper.

## Alternatives

- [Claude Code Ultimate Guide](https://ai-powered.com/mcp/claude-code-ultimate-guide): Claude Code Ultimate Guide is an open-source MCP server for ai models and llm tooling.
- [signafAIde](https://ai-powered.com/mcp/signafaide): signafAIde is a freemium MCP server for security.
- [Toolkien](https://ai-powered.com/mcp/toolkien): Toolkien is an open-source MCP server for ai models and llm tooling.
- [Lugh](https://ai-powered.com/mcp/lugh): Lugh is an open-source MCP server for ai models and llm tooling.
- [HypePaper](https://ai-powered.com/mcp/hypepaper): HypePaper is a freemium MCP server for ai models and llm tooling.
- [Agent Data Tools: Compact (find, describe, run)](https://ai-powered.com/mcp/agent-data-tools-compact-find-describe-run): Agent Data Tools: Compact (find, describe, run) is a freemium MCP server for developer tools.
- [Angel Agent Guard: pre-purchase check + prompt-injection scan](https://ai-powered.com/mcp/angel-agent-guard-pre-purchase-check-prompt-injection-scan): Angel Agent Guard: pre-purchase check + prompt-injection scan is a freemium MCP server for ai models and llm tooling.
- [Mizan Calibration](https://ai-powered.com/mcp/mizan-calibration): Mizan Calibration is a free MCP server for ai models and llm tooling.

---
Source: https://ai-powered.com/mcp/nerf-watch
