Skip to main content
Xerg records what was checked, where, when and with what coverage. It cannot guarantee that every production system is current or that an agent’s answer is correct. Observations keep their source and date. Scheduled graph updates determine whether those observations also update the editable graph.

Start with a serving domain

  1. Include the repository in the workspace’s GitHub setup. Connect Vercel from the graph (Accounts).
  2. Open Settings → Integrations → Evidence sources and choose Add source. An owner or admin can add a GitHub source or a Vercel source with its project, exact serving domain and environment.
  3. Choose how often to check it. Advanced sets the refresh interval, maximum evidence age and attempts per hour exactly.
  4. Each source is a row: its state, when it was last checked and how often. Select it to inspect the observed commit, deployment, routing and coverage before relying on it.
A source’s actions are in its … menu: refresh it now, edit it, disable it or rotate its signing secret. Anything that stops a source or revokes a secret asks first. GitHub reads are pinned to an immutable commit. Vercel evidence follows the serving domain, including reported rollouts; a successful preview build or a newer main branch does not prove what production serves. Periodic reconciliation recovers missed refresh hints. Failed access keeps the last observation with an unhealthy or stale state; it does not advance verification time. Only selected sources are covered. A domain observation cannot prove which revision served every request or region.

Audit a selected product journey

Open Settings → Integrations → Production checks and choose Add production check. This optional integration accepts a reviewed definition for one capability and audience. Select a restricted test identity, semantic navigation path, documentation URL and read-only evaluation endpoint that queries the actual bot. Known roles, plans, flags and regions need visible markers; unknown dimensions remain explicit. A runner on a host you control reads those surfaces independently. For your own product, Xerg checks the serving deployment before and after the run. Expired authentication is missing evidence, not proof a feature disappeared. Updating the expected path alone cannot correct outdated documentation or bot answers. Findings record evidence, affected version and owner. A review records the proposed correction; a later independent check must confirm resolution. Third-party checks cover only authorized surfaces, without claiming access to vendor internals. Collectors require an independently pinned definition and a source-only credential, created in Runner setup and advanced details, which expires in seven days. Browser cookies and source authorization stay on the runner. Browser journeys allow only same-origin GET/HEAD and block writes; select safe read-only endpoints and restricted identities. No automatic customer messages, purchases or correction publishing are included. Select View evidence on a change or problem to see the records behind that item, including the source, observation time and any missing detail. Recorded changes show their original facts; live assessments are labeled as current. Old Production context bookmarks open Evidence sources in Integrations. Production check results appear under Changes → Production checks. Problems is one list of what the checks and the graph found, with the checks still needing evidence under it. Select a problem to see its evidence and to review it: record a correction, acknowledge it, or snooze it with a reason and an expiry. Acknowledged, Snoozed and Resolved show the problems in those states. Viewers can read results; members can request a new check; owners and admins manage setup and finding decisions. Request new check queues work for your runner; it does not run a browser in Xerg. Schedule the runner separately and keep its credential current. A paused check stops new collection without clearing its findings. Missing or stale results never mean the check passed. Old audit-page bookmarks open Changes.

Runtime, contracts, webhooks and infrastructure

These checks are optional and limited to your own product and explicitly selected sources. Fresh bounded exports come from adapters on the trusted host. This release does not automatically connect every telemetry vendor, receiver or cloud service. Reports contain allowlisted metadata, source hashes and results; raw payloads and response bodies are discarded. See Security and data.

Require checks before a protected operation

Settings → Integrations → Protected operations defines the repository, environment, target, required sources, evidence policy and trusted command digest. Each protected operation is a row; Recent checks and reviews lists the checks made, with what each decided. Select a check to see why. One that needs an owner’s review is approved from there. The host that holds action credentials must run the configured gateway, which requests verification, consumes an exact short-lived single-use receipt and only then runs its fixed command. Missing, stale, contradictory or changed evidence blocks or requires the configured review. Authorized overrides record an owner and reason. The supported integrations are a trusted command wrapper and reusable CI workflow. They require operator installation and protected host configuration. Contact hello@xerg.ai for runner and gateway setup; these packages are not publicly distributed. For collector-backed policies, an active runner must complete a fresh collection within the 20-second verification wait or the operation cannot receive an ordinary passing receipt. Installing MCP alone does not force consultation. Commands run outside the configured gateway are not covered by its checks. Counts show received consultations, authorizations and host-reported outcomes over 90 days, not the percentage of all agent activity. External changes can still race after a check where the target lacks conditional execution. For customer-facing answers, render structured verified guidance when correctness matters; a prompt asking a model to consult Xerg is not an enforcement boundary.