@xerg/cli globally, replace npx @xerg/cli@latest with xerg.
Using a coding agent? Install the skill instead
If you work inside Claude Code, Codex, Cursor, or another agent that supports the Agent Skills standard, the fastest path is to install the Xerg skill and ask your agent to audit your AI spend:1. Run the first-run flow
init is interactive and walks you through your first audit. It:
- checks local OpenClaw, Hermes, and Claude Code defaults (current Hermes uses read-only
~/.hermes/state.db; Claude Code transcripts are read from~/.claude/projects/) - asks you to choose the runtime when more than one is detected
- runs the first audit and stores the local snapshot
- prints the normal terminal summary
- offers optional hosted follow-up after the audit succeeds
init prints the default paths it checked plus the next-step commands for:
- explicit local paths
- Claude Code transcripts elsewhere via
--claude-code-dir - Cursor usage CSV exports via
--cursor-usage-csv - event payloads from any framework via
npx @xerg/cli@latest ingest --file payload.json - optional OpenClaw trace collection via
npx @xerg/cli@latest collect openclaw; stop withCtrl-C, then review the local audit - optional certified Hermes trace enrichment via
npx @xerg/cli@latest collect hermes --state-db ~/.hermes/state.db; the first-party observer remains the recommended parity path npx @xerg/cli@latest audit --remote user@hostnpx @xerg/cli@latest audit --railway
2. Make one workflow or model change, then compare
--compare looks for the newest compatible cached snapshot on your machine and adds before and after deltas.
If no compatible baseline exists yet, Xerg does not fail. It adds a note telling you to run the same audit again after a fix.
3. Use direct commands when you want explicit control
Use the direct flows instead ofinit when you need non-interactive behavior, CI gates, or a specific output mode immediately:
--runtime openclaw, --runtime hermes, or --runtime claude-code.
4. Export or automate
Shareable Markdown:5. Optional hosted follow-up
If you want hosted features after the first local result, create a free workspace (or use an existing one):connectreuses existing auth when present, or starts browser login and offers to push the latest auditmcp-setupprints or writes hosted MCP config for Cursor, Claude Code, Codex, or another client (hosted MCP requires Pro or Enterprise)- the hosted dashboard shows pushed audits, sources, trends, Optimizations, policies, and workspace API keys
- the same workspace API key works for CLI push and hosted MCP
Common next steps
- Use init when you want the exact first-run behavior documented command-by-command.
- Use skill bundle and agent clients to set up the skill in your coding agent.
- Use audit to learn the full command surface.
- Use remote audits if OpenClaw is running on SSH or Railway.
- Use compare if you want to understand how snapshot compatibility works.
- Use waste taxonomy and metrics if you want the concepts behind the report.
- Use hosted dashboard, authentication and push, hosted MCP, or optimizations and policies if you want the hosted workflow after your first audit.