Xerg

Reviewed CLI version: 0.35.2.

The reviewed CLI pin is independent of the skill's release version. It selects an already-published, compatible package; a newer skill or package release does not itself authorize changing this pin.

Use this exact version for every skill-directed CLI action. The npm commands below select it explicitly; bare xerg examples are shorthand only for an explicitly approved installation whose xerg --version equals the reviewed version. Stop on a mismatch and request approval for the pinned installation; do not silently upgrade, use a mutable dist-tag, or fall back to another installed version. This pins the top-level CLI package, not its transitive dependencies or cryptographic integrity. Installation, local inspection, and hosted writes still require their separate permissions. Use browser pairing instead of requesting or displaying credential contents; the approved CLI manages its own authentication state.

Remote OpenClaw collection is available through the reviewed CLI 0.35.2, which includes the remote-command and archive security corrections. Obtain explicit permission for the exact host or Railway service and evidence paths before connecting or copying data. Use existing operator-managed authentication; never request credential contents in chat. Run the pinned doctor with the approved remote selection, verify its version, and preserve its shell-safe audit recommendation unchanged. Remote access does not authorize configuration changes or hosted upload. Prefer an approved local export when remote access is unavailable.

Xerg is a local-first CLI for finding wasted AI runtime spend. Runtime audits separate three evidence-strict monetary findings from seven neutral signals, report detector coverage, and use --compare to measure compatible changes. Runtime costs may be observed, locally estimated, or unpriced; they are not authoritative provider invoices.

Hosted pricing uses Monthly Audited Agent Spend from known spend in explicitly pushed runtime audits: Free through $2,000 per UTC month, Team at $99 through $10,000, Growth at $299 through $50,000, Scale at $799 through $250,000, and custom Enterprise above $250,000. Team, Growth, and Scale have identical capabilities. Historical imports, duplicate-review spend, and unpriced usage do not count toward MAAS.

Cold-fetch setup

When this file was fetched from https://xerg.ai/skill.md (including via the link on the npm listing for @xerg/cli) rather than loaded from an installed skill:

1. Identify the current agent host. OpenClaw, Hermes, QM, Claude Code, Cursor, and Codex are setup hosts; they are not automatically the data source being audited. QM can also be the audit source, but only through its explicit administrator-provisioned Xerg connection.
2. Check whether Xerg is already installed for this host. Before writing to a persistent or global skill directory, explain the destination and ask the user for permission.
3. If the host has a supported skill installer, use it after permission. OpenClaw can install xerg from ClawHub. For Agent Skills installation, require the operator to supply a reviewed 40-character hexadecimal commit SHA from the public xergai/skills repository. After explicit permission for the installer download and skill installation, use npx --yes --package=skills@1.5.25 skills add 'https://github.com/xergai/skills/tree/<REVIEWED_PUBLIC_REPO_FULL_SHA>/skills/xerg' --skill xerg, replacing only the placeholder with that SHA. Do not execute the placeholder, use a branch or tag, discover a latest revision automatically, or add unattended agent/scope selections. The pinned installer requires Node 22.20 or newer; do not upgrade Node implicitly. Without a reviewed SHA, stop the installer step or use a separately approved, already-reviewed local skill folder through the host's supported import flow. You may still follow this fetched file for the current session under the separate permissions in steps 4–5; do not persist it implicitly. The reviewed CLI version does not prove an older bundled skill is current. Hermes can use Xerg's well-known skill discovery. Deterministic host-specific commands live at xerg.ai/agents.
4. If persistent installation is unavailable or declined, use these fetched instructions for the current session. Do not invent an install location.
5. Ask permission before any npm download or install, including the transient package fetch that npx may perform. Obtain separate permission to inspect local runtime data. After both applicable approvals, run npx --yes @xerg/cli@0.35.2 doctor --json, parse its cliVersion, canAudit, and recommendedCommand, and apply the version-continuity check below before offering the audit. On QM, do not use bare doctor: follow the private-scope QM procedure below.

Always ask before uploading. A local audit needs no account and must not be pushed automatically. On an approved hosted command only, set XERG_SETUP_METHOD=skill and set XERG_AGENT_HOST to the current execution host: claude_code, cursor, hermes, codex, openclaw, qm, or other. These are best-effort, content-free analytics metadata, not authentication or the audit source. If the user's setup prompt includes an Xerg organization ID, preserve it with --organization-id on activation.

After a successful local result, keep hosted work in separately approved stages. First offer an optional local payload preview. It prints the intended payload, so summarize its included fields plus the exact auditId and generatedAt without pasting the full payload into chat. The default preview uses the newest snapshot in the default database; use push --file <path> --dry-run for an explicitly selected prepared payload.

XERG_SETUP_METHOD=skill XERG_AGENT_HOST=codex npx --yes @xerg/cli@0.35.2 push --dry-run

After separate approval to connect, run XERG_SETUP_METHOD=skill XERG_AGENT_HOST=<host> npx --yes @xerg/cli@0.35.2 activate --connect-only. Add --organization-id <verified-org-id> only when the user supplied or verified that exact value; do not execute placeholders. activate opens Xerg in the browser. The signed-in user explicitly approves the currently active workspace; the page shows its full organization ID, live plan, and API environment before any credential is issued. The workspace key is encrypted to the initiating CLI and stored with owner-only, environment-bound metadata. Use browser pairing instead of requesting a pasted key or exposing one in a command.

In skill-directed work, connection must stop after verified credential storage and must not be combined with an audit or push flag. After another explicit upload approval, run XERG_SETUP_METHOD=skill XERG_AGENT_HOST=<host> npx --yes @xerg/cli@0.35.2 push, adding --file <path> only when the preview used that same validated payload. Human-operated CLI flows may still use other documented activation modes.

Codex is an execution host, not a native Xerg audit source. A Codex user audits whatever doctor finds: OpenClaw, Hermes, Claude Code, a Cursor export, or a generic ingest payload.

Local commands remain telemetry-free and add no analytics call. Pairing creation and explicit push requests send a fixed, content-free execution-context envelope. Do not infer Codex from CODEX_*, process names, ancestry, or undocumented runtime state; use the explicit hint above only when this skill is running the user-approved hosted command.

If the current host is QM, a cold-fetch request can explain prerequisites but cannot claim it persisted an admin skill pack, created PostgreSQL privileges, or provisioned secrets. Require an administrator-only private scope. Refuse deployment-wide output in public/shared scopes. Direct users to approved secret-entry mechanisms instead of requesting a DSN, identity key, Fly token, provider credential, or Xerg API key in Slack or chat.

First run (agent execution path)

Use the non-interactive commands. Do not run xerg init unless the user explicitly asks for guided interactive setup — it is TTY-only and prompts for input.

Before running the command, ask the user for permission for any npm download or install and separately for permission to inspect local runtime data. A persistent skill installation, a transient package fetch, and local analysis are distinct actions; do not infer approval for one from another.

npx --yes @xerg/cli@0.35.2 doctor --json

1. Run doctor in JSON mode first. It returns the CLI version, runtime/product, canAudit, checked sources, runtime evidence, notes, pricing-catalog provenance, and a shell-safe recommendedCommand. Check that its CLI version equals the reviewed version before continuing. QM is never auto-probed.
2. If doctor resolves Hermes, classify the purpose before starting any new Xerg-directed workload. Always run npx --yes @xerg/cli@0.35.2 doctor --runtime hermes --require-observer-live --json first. A workload intended to test observer-backed or sequence-dependent findings must hard-stop when this preflight exits 5. For an already-existing or historical aggregate-economics audit, show the aggregate-only coverage warning and continue with the ordinary recommended audit command; live observer health is not a prerequisite for reading existing economics. Treat this as a continuous production-coverage check, not merely a test prerequisite.
3. When canAudit is true, require the recommendation to select reviewed CLI version 0.35.2. Preserve its exact runtime, path, flags, shell quoting, and argument order; show it and obtain audit permission before running it exactly as printed. Stop if the recommendation selects another version or a mutable package reference. A bare xerg recommendation requires the version-verified installation above.
4. If an old global CLI or stale snapshot emits a finding with ruleId: tool_sequence_repetition_v1, do not summarize or recommend from that amount. Rerun doctor and the audit through npx --yes @xerg/cli@0.35.2, then use only the current result.
5. When any active-finding coverage is none, partial, unsupported, or unknown, put the aggregate-only coverage warning before all spend totals and conclusions. State that aggregate economics remain available where priced, identify the unavailable sequence-dependent analysis, explain that missing historical request sequence cannot be reconstructed and repairs cover only future activity, and never translate zero identified findings into “no waste.” Lead with pricing coverage whenever any calls/tokens are unpriced or limited-estimate. Then summarize runtime spend (or known runtime spend when pricing is partial), identified waste (wasteSpendUsd), assessed spend from detectionCoverage, active findings, neutral signals, and spendByAgent when present. Delegation rows show spend that is already counted in the child row and the audit total; never add them again.
6. If an OpenClaw result says session transcripts were authoritative over non-empty gateway evidence, repeat that limitation before totals and conclusions. State that gateway-only activity may be omitted, activity and spend may be understated, and findings may differ from a fully reconciled view. Do not describe identified waste as a lower bound or treat a passing monetary/detector-coverage gate as certification of omitted gateway-only activity. Use only --sessions-dir or only --log-file when the user explicitly requests one local source kind; never add the excluded gateway count back into the transcript result.
7. Before recommending remediation for any monetary finding, pin the lookup to the audit JSON that produced it: run npx --yes @xerg/cli@0.35.2 explain <finding-id> --audit <auditId>:<generatedAt> --db <dbPath> --json, using the JSON's exact auditId, generatedAt, and dbPath values with shell-safe quoting. Never use a bare newest-snapshot lookup. If dbPath is absent because the audit used --no-db, inspect its inline evidence packet and explain that later lookup is unavailable. Verify the required observations, economic meaning, pricing coverage, limitations, and next step; do not recommend a fix from the summary row alone.
8. Present repeated tool-chain and deep-loop-activity rows as neutral observations. A high distinct ratio is consistent with fan-out and argues against an identical-input loop, but does not prove success. A low ratio is a reason to inspect, not proof of waste or recoverability. Associated spend is not classified as waste.
9. If the user applies a fix, re-run the same audit with --compare to report the before/after delta:

npx --yes @xerg/cli@0.35.2 audit --json --compare

Report monetary before/after values only when comparison.spendComparisonAvailable is explicitly true; a missing flag cannot establish baseline pricing and is unavailable. Report identified-waste deltas only when both spendComparisonAvailable and wasteComparisonAvailable are true.

If no local data is found, doctor prints the paths it checked. Fallbacks:

• Cursor usage CSV export: npx --yes @xerg/cli@0.35.2 audit --cursor-usage-csv ./cursor-usage.csv
• Claude Code transcripts elsewhere: --claude-code-dir <path>
• Hermes profile database: --runtime hermes --state-db <path>
• Any framework's exported event payload: npx --yes @xerg/cli@0.35.2 ingest --file payload.json
• Existing sanitized OpenClaw trace capture: npx --yes @xerg/cli@0.35.2 audit --otlp-file <capture.jsonl>
• New local OpenClaw trace capture: npx --yes @xerg/cli@0.35.2 collect openclaw (interactive until Ctrl-C; use only when the user asks to collect a workload)
• Certified local Hermes trace enrichment: npx --yes @xerg/cli@0.35.2 collect hermes --state-db <path> (interactive until Ctrl-C; state.db remains required)
• Remote OpenClaw evidence: after permission for the host and evidence paths, use pinned doctor with --remote <host> or --railway and the approved remote path flags; never execute placeholders or infer an unapproved target
• OpenClaw evidence exported from another host: place the approved export on the local machine, then select it with a local --log-file, --sessions-dir, --openclaw-db, or --otlp-file path
• Existing QM snapshot: npx --yes @xerg/cli@0.35.2 audit --runtime qm --qm-snapshot <snapshot.jsonl>
• Configured QM direct/Fly source: an operator collects outside Slack; follow the private-scope procedure below only for an authorized snapshot

Use the npx --yes @xerg/cli@0.35.2 path for first-run and rerun analysis so an old global CLI cannot reintroduce the retired name-only detector. A user may choose an explicitly approved global install verified against this skill's reviewed version. A newer-version notice is informational, not permission to change the pin; update the skill and approve its reviewed CLI version before using that version for skill-directed work.

The reviewed CLI requires Node 22.15.0 or newer; Node 23 requires 23.8.0 or newer. Do not upgrade the operator's Node installation implicitly.

What It Audits

• OpenClaw database snapshots, gateway logs and session transcripts; when both produce calls in the requested window, transcripts are the sole audit authority and gateway runs are excluded
• Optional independent OpenClaw trace captures created by the loopback traces-only collector
• Hermes v0.17-v0.20.1 state.db (read-only), with legacy log/transcript fallback where present
• QM durable run/model economics and current retained activity through xerg_export/v1 and qm-snapshot/v1
• Claude Code session transcripts via xerg audit --runtime claude-code
• Cursor usage CSV exports via xerg audit --cursor-usage-csv ./cursor-usage.csv
• Any framework's exported event payload via xerg ingest --file payload.json

Xerg does not currently ingest provider bills, reconcile invoices, or convert runtime observations into FOCUS. If a user asks for one of those capabilities, explain the boundary and do not present modeled runtime spend as invoice-authoritative.

What It Finds

• Retry waste only from a stable charged failed/aborted attempt chain that ends in a higher successful attempt
• Tool-loop waste only from exact repeated tool name/input/result/state evidence with no progress and exact cost correlation
• Cache churn only when a cache-entry lifecycle costs more than its uncached counterfactual
• Neutral signals for deep loops, context outliers/growth, fixed cadence, premium-model routine labels, cache-read concentration, and Max Mode concentration; ordered metrics, optional human-readable workflow/run labels, and optional associated spend remain descriptive, and signals have no avoidable-spend, recommendation, optimization, or CI effect
• Local digest-only evidence packets and xerg explain for findings, signals, and qualifying repeated tool chains; raw and truncated tool arguments/results are never retained
• Claude Code streaming reconstruction and argument-aware chain diagnostics; Claude remains ineligible for monetary tool-loop findings because its transcript supplies no defensible state/progress fingerprint
• Per-agent spend attribution, including observed immediate parent-to-child delegation paths for Claude Code sidechains and ingest payloads. Delegation rows show spend that is already counted in the child row and the audit total; never add them again.
• Cost per outcome when runs carry outcome signals; declare outcomes with xerg outcome --workflow <name> --status success|failure
• Separate local Hermes mechanical metrics when the optional observer is enabled; these have no dollar classification, recommendation impact, or CI-gate effect
• Shared local analysisCoverage, toolActivity, and workloadEconomics blocks for OpenClaw and Hermes; these are neutral evidence and never affect findings, recommendations, waste totals, or CI gates
• QM request-versus-aggregate reconciliation, pricing coverage, source stability, pseudonymous scope attribution, and current-window unassociated tool activity

Costs are priced across input, output, cache read, and cache write tokens using a reviewed local catalog covering hundreds of current models. Opus 5 includes separate five-minute and one-hour cache-write rates. If a transcript preserves only aggregate cache creation, Xerg uses the documented five-minute default, reports affected tokens, and discloses the possible additional one-hour-cache cost. Audits never fetch pricing from the network. Run xerg doctor --verbose to see per-file extraction coverage and any files excluded by OpenClaw source authority.

For current Hermes, use xerg audit --runtime hermes; Xerg prefers $HERMES_HOME/state.db, where HERMES_HOME is the selected Hermes profile directory. Xerg 0.24.2 supports Hermes v0.17-v0.20.1 and certifies state schemas 16-25; later schemas are best-effort and explicitly uncertified. Use --state-db for another profile. --state-db is mutually exclusive with legacy --log-file and --sessions-dir. State-only audits retain aggregate spend, request/token/cache, workflow/model/tool, and delegated-workload totals, but first-request cost, initial context, request growth, retry sequences, and identical-input loops are unavailable. A Hermes operator may separately install the optional xergai/hermes-observer plugin described in Hermes analysis. When present, Xerg reads its content-free local events and bounded health state. This skill does not install, configure, start, restart, or modify observer state. Complete observer evidence can split an aggregate only after exact request/token reconciliation; observer records never enter economics or audit identity.

Hermes v0.20.1 can record auxiliary tasks such as title_generation in state.db without firing its public per-request plugin hooks. Preserve their economics as task-scoped aggregates, disclose that their sequence-dependent analysis is unavailable, and do not let them invalidate exact evidence for observable main/delegated requests. Never call a mixed result full request coverage or imply that the observer transparently instruments Hermes's private auxiliary client.

On Hermes v0.20.x, treat generated and truncated terminal quantities labeled “At least” as conservative UTF-8 byte floors; returned bytes remain exact. Do not difference floors, even against other floors, and do not infer or compare generated/truncated values when the measurement basis is missing. Xerg never reads or retains full_output_path. A local source-integrity warning that says totals may be floors means Hermes's usage primary key may have dropped distinct task rows; also state the narrower guarantee that totals are exact for rows Hermes recorded. The warning is advisory, non-repairing, local-only, outside audit identity, and absent from pushed payloads. Xerg 0.24.0 was not certified for Hermes v0.20.x terminal mechanics and could understate those byte values; 0.24.1 is the fix.

After the operator installs the observer and restarts Hermes, run xerg doctor --runtime hermes --require-observer-live before starting new Xerg-directed activity. The strict preflight succeeds only when a current observer process is running and its writer is not known unhealthy; otherwise it exits 5 with the reason. A restart can cover only future activity and never reconstructs historical aggregate sessions. Ordinary historical audits remain available, but their aggregate-only limitation must be disclosed before totals or conclusions. Use --require-detection-coverage full|partial for audit-result CI; unmet coverage also exits 5.

For OpenClaw traces, xerg collect openclaw binds only to 127.0.0.1, accepts OTLP/HTTP protobuf traces, persists a bounded sanitized capture, and audits after shutdown. It does not modify OpenClaw configuration or push automatically. Without explicit --runtime hermes, --otlp-file is independent OpenClaw evidence and cannot be combined with transcript, log, other-runtime, or remote sources.

For certified Hermes traces, xerg collect hermes uses the same loopback traces-only protocol but HMACs identifiers with a persistent local key and requires state.db. Use only the exact certified briancaffey/hermes-otel commit printed in Xerg's docs. Xerg never installs or edits that plugin. The first-party observer stays primary; cache-inclusive plugin prompt totals are normalized into Xerg's separate input/cache buckets, optional enrichment preserves economicAuditId while analysis identity reflects changed coverage, and conflicting evidence restores the original aggregate. For non-default HERMES_HOME profiles, use the equivalent environment-variable block printed by the command because the pinned plugin resolves YAML from the selected Hermes profile directory.

QM private-scope procedure

QM currently supports bounded one-shot collection through a host-independent snapshot adapter. An operator outside Slack creates qm-snapshot/v1 through either a strict view-only direct reader or the Fly-contained one-shot exporter. Fly contained mode uses QM core's existing DATABASE_URL only inside the hidden exporter, records a process boundary with no database-level least-privilege claim, and never exposes that credential to this skill. Do not mutate setup, install/persist the skill, activate, or push without explicit approval.

Offline QM Slack audits accept only an authorized pre-created snapshot. The CLI must be explicitly approved, provisioned, and version-verified by an operator in the administrator's private runtime; agents must not install it implicitly or initiate live collection. Verify xerg --version against this skill's reviewed version before doctor; on a mismatch stop and ask the operator to provision that version, without an npm fetch from Slack. The runtime receives no database URL, identity key, Fly token, provider credential, or Xerg credential. Continuous follow capture, durable tool-history capture beyond QM's retention window, and live Slack-triggered collection are not currently supported.

Run doctor before every first audit:

xerg doctor --runtime qm --qm-snapshot <snapshot.jsonl>
xerg audit --runtime qm --qm-snapshot <snapshot.jsonl> --since 7d --json

Operator documentation may explain --database-url-env, --fly-app, --print-sql, and --print-views-sql, but the agent must not execute or provision those live paths from Slack. Direct, Fly, and --qm-snapshot modes are mutually exclusive.

Explain the result conservatively:

• positive QM cost is observed; an explicit deterministic model may be catalog-estimated
• zero/negative placeholders and unresolved openrouter/auto are unpriced, never free or actual $0
• unpriced patterns say “monetary impact unavailable” and monetary gates exit 5
• Pi/OpenCode timestamps are flush/capture time; Claude is near result-step recording; Codex rows are turn aggregates
• request-sequence detectors run only on exactly reconciled request observations; aggregate rows are excluded
• QM keeps activity for one hour, so older tool history is unavailable and tool-to-model association is not claimed
• retry attempts are neutral evidence without spend attribution; unsupported findings are not described as assessed

Local snapshot audit is the default. Ask for explicit approval before xerg activate --runtime qm, audit --push, or any other hosted write. Xerg's hosted MCP can read an already-pushed pseudonymous summary but has no native QM database or Fly access.

Optional Cloud

Local audits need no account. To connect a workspace and push the latest local result, ask permission and run:

After selecting the exact bounded value from the mapping above, put it on the same approved hosted command. Codex example:

XERG_SETUP_METHOD=skill XERG_AGENT_HOST=codex npx --yes @xerg/cli@0.35.2 activate --connect-only

For a website-first user with no cached audit, first complete the separately approved local audit. Pair with activate --connect-only, optionally inspect push --dry-run, and request a distinct approval before push. Skill-directed setup does not combine these stages. Hosted sync and hosted MCP remain optional and never run without explicit user action.

After an approved push, the hosted Fleet Overview leads with strictly validated seven-day before/after savings when available, keeps current waste separate, can group identified-waste rates by department or runtime posture, and retains Source Performance. Department and runtime-posture assignments are current attributes per source on Free and paid plans. Runtime posture remains Not set until a workspace admin assigns one of vpc, on-prem, saas, local, hybrid, or air-gap; Xerg never infers it from framework, host, collection environment, or audit evidence. Audit Detail adds daily Performance and detector-attributed Workflow economics; Compare retains outcome evidence, and every Signal presentation keeps suggested investigation guidance explicitly separate from savings estimates.

When a user supplies a hosted Optimization ID and has already authorized hosted MCP access, retrieve its current safe handoff with xerg_get_optimization({ optimization_id }) before proposing a code change. Treat the handoff as scoped evidence and instructions, inspect the target repository independently, make no automatic policy change, and require a compare audit for validation. A qualified validation fact may be described as qualified savings; an unqualified fact must retain its reason, and a missing fact is only “Validation details unavailable.”

If a user asks to send an Optimization to Linear, use the hosted dashboard action only after a workspace admin has connected Linear and selected a team. Explain that Xerg creates one sanitized, one-way issue from the existing handoff; it does not send raw model or tool content, synchronize issue status, or change the Optimization lifecycle automatically. Do not claim an issue was created unless the dashboard returns its stored Linear link.

Advanced authentication

Use xerg login --replace only as a manual recovery path when browser pairing is unavailable. It opens Workspace Settings and masks terminal input. Direct the user to that terminal flow rather than requesting the workspace key in agent chat.

Use XERG_API_KEY only for non-interactive CI or deployment automation. Store it in the CI provider's secret manager; never place it inline in a shell command, source file, log, URL, or conversation.

Links

• Docs: xerg.ai/docs
• Service status: status.xerg.ai
• Skill: xerg.ai/skill.md
• npm: @xerg/cli
• Support: hello@xerg.ai