AgentDX records what your AI coding agents actually do — tokens burned, models called, tools fired, costs incurred. Across Claude Code, Cursor, Codex, Aider, and Cline. All on your machine. None of it leaves.
Zero config. Detects your agents automatically. Free forever.
No agents to install, no accounts to create, no data to upload. AgentDX reads the session logs your tools already write to disk.
AgentDX scans for local session data from every supported agent — no setup, no config files.
$ npx agentdx
Every session is parsed into a local SQLite ledger: tokens, models, tools, timings, errors.
→ 1,247 sessions recorded
A local dashboard opens in your browser. Inspect, filter, drill down to any single turn.
→ localhost:3002
When something works — or doesn't — you have the full record.
Forensic timeline of every message, tool call, and token spend. Reconstruct exactly what the agent did, in order, with timing.
Input, output, cache reads, cache writes — broken down per model, per session, per repo.
See which models you reach for, how they perform, and what they cost you over time.
No code, prompts, file paths, or repo names. Ever. AgentDX records the shape of your work, not the substance.
One SQLite file in your home directory. No cloud, no account, no telemetry you didn't opt into.
~/.agentdx/agentdx.db
MIT licensed, end to end. Audit the collectors, extend them, add your own agent.
github.com/agentdxhq/agentdx ↗Devs run AgentDX locally. Managers and CTOs see the aggregate — anonymous, opt-in, never the code. Cost, adoption, and usage patterns across the whole org.
One command. It finds your agents and opens the dashboard.
Requires Node 18+. Reads from ~/.claude, ~/.codex, ~/.cursor, and more.