Agent Profiler

Agent Profiler is a local-first tool for understanding how AI coding agents spend observable effort in your workspace. It records Cursor and Codex hook traffic into SQLite so you can review session shape, estimated token usage, tool and shell noise, always-on context weight, and simple efficiency signals—without sending telemetry to a remote service.

Today the workflow is intentionally scoped project-by-project: you run commands from a repo root, store profiler config and data under .agent-profiler/ beside that project, and inspect sessions with the CLI or the optional dashboard. A future direction is to offer a clearer opt-in path to operate primarily from the home-directory profile (for example global installs and multi-repo DB layout) when users want that; the current release optimizes for per-repo isolation and predictable paths.

On agent-profiler init in dev mode, when the profiler directory resolves to <project>/.agent-profiler, the tool appends a .gitignore rule so the local store (including SQLite and synced dashboard assets) is not committed by mistake.

Dashboard (early version)

Run a minimal localhost-only web UI over the same database and context audit used by last:

agent-profiler dashboard
# open http://127.0.0.1:3737/ — use Refresh to reload metrics

The dashboard (early version) surfaces:

• Observable usage — estimated input, output, tool/MCP result, and shell output tokens with simple bar proportions
• Efficiency score — heuristic score with a small sparkline over recent keyed sessions (when present)
• Session timeline — chronological strip colored by role (user, assistant, tool, shell, other)
• Tool result sizes — histogram buckets for large tool payloads
• Context audit — estimated tokens for common always-on instruction paths in the repo
• Red flags and recommendations — aligned with the last command logic

Static assets are copied into .agent-profiler/dashboard/ when the server starts so the UI stays next to your project data. Expect rough edges and UI iteration in future releases.

Requirements

• Node.js 22 or newer
• macOS, Linux, or another environment supported by better-sqlite3

Install

For one-off commands, use npx:

npx agent-profiler --help
npx agent-profiler last

For persistent hook installation, install the package so agent-profiler is available on PATH:

npm install -g agent-profiler
agent-profiler --help

Quick Start

Initialize hooks for Cursor:

agent-profiler init cursor --mode prod

Initialize hooks for Codex:

agent-profiler init codex --mode prod

Inspect the resulting setup and the latest captured session:

agent-profiler status
agent-profiler last
agent-profiler audit context

Hook approval and restarts

• Codex: Hooks must be approved before they run—confirm when prompted or enable them in the Codex plugin settings.
• Cursor: Restart Cursor after init so hook configuration reliably takes effect.

npx vs --mode prod

• npx agent-profiler ... is great for one-off inspection commands.
• agent-profiler init <source> --mode prod writes persistent hook commands that expect agent-profiler to be installed on PATH.
• --mode dev is meant for local development in this repository and writes absolute node <repo>/dist/cli.js ... hook commands instead.

Commands

• agent-profiler init <cursor|codex>: install hook wiring for a supported source
• agent-profiler hook <source> <eventName>: ingest one hook payload from stdin
• agent-profiler status: inspect local setup and ingest state
• agent-profiler last: summarize the most recent observed session
• agent-profiler dashboard: serve the local dashboard (SQLite + context audit)
• agent-profiler audit context: estimate always-on context token footprint

Releases

Releases are automated with semantic-release. Pull requests run CI plus canary publishing, and pushes to main publish to npm and create a GitHub release.