A Rust CLI & TUI that shows what your coding agent loads and runs — secrets it can read, always-loaded instructions, MCP tools per agent, skills and shell overhead — and fixes what it can. Every number is measured, never guessed.
agentprof fix writes Read deny rules — Claude Code ignores .claudeignoreagentprof bench measures yoursWritten in Rust. Every figure it reports is measured, never guessed.
Claude Code does not read .claudeignore; only Read(...) deny rules in .claude/settings.json keep files away from it. agentprof checks every .env, key and credential file against the deny rules actually in effect, and agentprof fix merges the missing rules in — keeping your settings and a backup.
Performs a real initialize → tools/list handshake with every configured server and reports the load per agent — full definitions where they are loaded up front, tool names only where Claude Code defers them. Servers defined inside a repository are only started with --trust-workspace.
Agents keep every skill's name and description in context and load SKILL.md only when it runs. agentprof skills reports both, flags oversized skills and missing descriptions, and finds skills competing for the same intent (review, qa, design).
Finds instruction files that prescribe conflicting conventions (e.g. ObservableObject vs @Observable, npm vs pnpm) while ignoring rejections like "instead of". agentprof compile moves language-specific sections into path-scoped rules — .claude/rules, .cursor/rules, Copilot instructions — so they load only when relevant.
uses: dautovri/agentprof@v0.3.0 installs a checksum-verified binary, writes the health report to the job summary, and fails the pull request when the score drops, a secret becomes readable by Claude Code, or instruction files contradict each other. Scores use only repository contents, so every runner agrees. SARIF output feeds GitHub code scanning.
Estimate the overhead agentprof surfaces across a team, so you can decide what is worth cutting.
agentprof bench and agentprof mcp --probe for your own numbers.