A Rust CLI & TUI that measures the hidden cost of every agent turn — subshell startup latency, MCP tool-schema tokens, instruction bloat, and skill collisions. Every number is measured, never guessed.
Written in Rust. Every figure it reports is measured, never guessed.
Every AI agent tool execution forks a subshell. If your ~/.zshrc runs slow version managers (nvm, pyenv) or git status prompts, 50 tool executions waste over 30 seconds. agentprof injects an idempotent, zero-overhead bypass header so agent subshells run in under 6ms.
Connected Model Context Protocol servers inject extensive JSON tool definitions into your prompt on every turn. agentprof mcp profiles schema token overhead and flags heavyweight tools before they blow through your budget.
Audits all installed skills across Claude Code, OpenCode, and local worktrees. Identifies keyword overlaps where multiple skills compete for identical intents (e.g. review, qa, design) and cause hallucination loops.
Detects conflicting rules across AGENTS.md, CLAUDE.md, and .cursorrules (e.g. ObservableObject vs @Observable, deployment target mismatches). Compresses bloated rule files and compiles them into context-routed JIT instructions.
Prevents AI agent search and grep tools from indexing massive unignored directories (.build/, DerivedData/, node_modules/, target/) or accidentally ingesting sensitive credential files (.env, *.pem, credentials.json). Merges patterns into .claudeignore and .cursorignore without discarding existing rules, keeping a backup.
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.