MCPHawk 1.0: DevTools for MCP

• One command setup: mcphawk install routes every MCP server of Claude Desktop, Claude Code, Cursor and VS Code through MCPHawk, with a backup of each config and a clean mcphawk uninstall
• Agent runs: everything a client does across all its MCP servers on one waterfall timeline, split into runs by task, with status, latency and result size for every call
• Call inspector: rendered tool results, full request and response, headers and a shareable link for every call
• Replay: edit any call and send it again to the real server, recorded next to the original for comparison
• Context cost: tokens each server and tool adds to every turn, tools that never get used and the results that flood the context window
• Problems: JSON-RPC errors, tool failures, unanswered and slow calls, agent loops and spec violations, most severe first
• MCP 2026-07-28 support: stateless requests, multi round-trip requests, server/discover, Mcp-Method headers and cache hints, alongside the classic handshake era, with spec lint for the version each session speaks
• Compare: tools added, removed or changed between two versions of a server, including the token delta, plus how each call behaved
• MCP server for agents: seven read-only tools let Claude Code list runs, find failing calls and context hogs, and link you straight to them in the UI
• OpenTelemetry export: stream metrics, logs and traces over OTLP to Grafana, Datadog or any backend, following the OpenTelemetry conventions for MCP
• Prometheus and Grafana: a /metrics endpoint and a ready made Grafana dashboard for requests, failures, latency and context cost
• HTTP proxy: record Streamable HTTP and HTTP+SSE servers, local or remote, including HTTPS
• Passive sniffing: capture plaintext local MCP traffic with zero config changes
• Private by default: a local database, secrets masked before they are stored and a UI that only listens on localhost
• New web UI: a DevTools style interface with setup guidance, live updates and deep links for every run and call