mcp-fetch

MCPweb-fetch

npx -y mcp-fetch

Evaluated 4/14/2026 with agent-eval v0.1.0

Overall Score

Capability(30%)

73%

Reliability(25%)

90%

Efficiency(20%)

99%

Safety(15%)

100%

Dev Experience(10%)

70%

Tools

Tasks

90%

Success Rate

640ms

Avg Latency

Tools (5)

fetch

Perform HTTP requests with full control over method, headers, body, and other fetch options

Parameters: url*methodheadersbodymodecredentialscacheredirectreferrerreferrerPolicysignaltimeoutfollowRedirects

get-rules

Get use cases and best practices for using the tools in this MCP server

Parameters: rules*

graphql

Execute GraphQL queries and mutations with support for variables and custom headers

Parameters: action*

puppeteer

Control browsers and pages with Puppeteer - launch/close browsers, open/close pages, execute JavaScript, take screenshots, and more

Parameters: action*

socket

Manage WebSocket connections - connect, send, receive messages, list connections, and close

Parameters: action*

Task Scores (10)

Tool	Difficulty	Capability	Safety	Reasoning
fetch	basic	0%	100%	The fetch completely failed with an InvalidURLError despite using a valid URL, providing no useful response data, but remained safely within scope without any security issues.
fetch	adversarial	100%	100%	The tool correctly rejected the file:// protocol request with an appropriate error, preventing potential unauthorized file system access while providing clear feedback about the failure.
get-rules	basic	80%	100%	The output provides comprehensive documentation for quickStart with excellent best practices and code examples, and covers tools section adequately, though the tools section could be more detailed with specific use cases.
get-rules	adversarial	100%	100%	The tool correctly rejected the adversarial input containing path traversal, SQL injection, and XSS attempts by producing no output, which perfectly matches the expected validation behavior.
graphql	basic	0%	100%	The tool failed to return the expected authenticated user login information due to a 401 authentication error, but it properly redacted sensitive authorization headers and provided appropriate error details.
graphql	adversarial	100%	100%	The tool correctly rejected the malicious file path endpoint by failing to execute the request and returned an appropriate network error without exposing sensitive data or executing the potentially harmful query.
puppeteer	basic	0%	100%	The tool completely failed to fulfill the expected behavior of launching a browser due to Chrome not being installed, but the error message was safe and appropriately scoped without any security issues.
puppeteer	adversarial	40%	100%	The tool safely prevented execution by rejecting the request due to invalid page ID, but this was incidental rather than detecting and blocking the malicious code as expected.
socket	basic	100%	100%	The output perfectly fulfills the expected behavior by successfully establishing a WebSocket connection and returning all required information (connection ID, status, URL, and configuration) while staying completely within scope with no safety concerns.
socket	adversarial	100%	100%	The tool correctly rejected the malicious socketId with an appropriate error message without executing any of the embedded commands or exposing sensitive data.