Security
Your code
ZPX Forge does not store code. Agents work on each developer's machine; the platform only keeps metadata (projects, tasks, events) and what GitHub exposes to the app you install.
Isolated companies
All business data belongs to a company and is queried through it. Anyone who is not a member gets not found. Automated tests verify this on every change.
Accounts
Web sign-in requires GitHub and a verified email. To accept an invitation or decide on a release, use the account associated with the designated email. Forge keeps the session in a protected cookie and checks the origin of requests that change data.
GitHub
- An installation is only linked if the person controls the GitHub account.
- Webhooks are verified with their signature and processed only once.
- The app's private key never leaves the server.
Infrastructure
Browser → zpxforge.com (Vercel) → AWS load balancer (HTTPS) → private backend
The backend has no public IP and only accepts traffic that went through Vercel, identified with a shared key.
Audit log
Every relevant action (registration, approvals, invitations, roles, repos, GitHub) is recorded with who, when and from where.
Shared sheets
Anyone with the public link can view the snapshot for 30 days or until revocation. To decide on a delivery, an external customer opens the private link using the verified email account designated for the milestone.
Pro security audit
From a linked target repository, sign in with zpx login, select your project with zpx use company/project, then run zpx audit. Use CLI 0.3.0 for Pro. Use zpx audit --pdf report.pdf to save the PDF locally when recording a revision. The basic zpx audit still works without an account. Pro reads committed files, checks Python, JavaScript and TypeScript source, flags environment files and missing workflows, detects npm deprecations and looks up public npm and PyPI versions from package-lock.json and uv.lock in OSV. It compares direct public dependencies with registry releases and checks Kubernetes memory limits and readiness probes. It also checks synchronous requests without explicit timeouts and dynamic privileged messages. For a repository-owned JSON message template ending in .prompt.json, it inserts an inert sample at {{document}} and checks which message role receives it. The report records coverage. Runtime load and model behavior require separate evaluation.
Add --agent claude or --agent codex to generate an executive interpretation with the Forge security auditor profile. The selected client receives totals, up to 15 prioritized findings and up to 10 coverage gaps. Its attributed commentary appears in the portal and PDF alongside severity, evidence and remediation tables. The responsible reviewer retains release authority.
Open Project → Auditoría de seguridad for the private versioned PDF. Each finding generates a proposed specification. The spec owner can select individual findings or use Seleccionar todas. Approving the selection adds new tasks to the earliest open wave. Adding tasks to an approved or active wave also requires plan approval rights. If all waves are closed, the plan owner must explicitly choose where to continue. After committing fixes, run the command again to compare new, persistent and no-longer-detected findings. Incomplete coverage leaves earlier findings unverified. A release decision owner must approve the latest report for the exact commit. High, critical and unclassified risks and incomplete checks block approval.
In Plan → Gestión del plan, choose Revisiones de seguridad to filter historical remediation by status. Open Entregas → Auditoría de seguridad for the audit report history.
For actual deployment enforcement, enable the ZPX Forge custom protection rule on the GitHub production environment used by the deployment job. Grant the installation Deployments write permission and subscribe the app to the deployment_protection_rule event. After rejection, resolve findings, audit and approve the new revision, and rerun the workflow.
The same command checks your subscription and project permissions. Pro enables the full report automatically. Individual runs the basic diagnostic. Requesting a PDF or AI interpretation without access returns an explanation that Pro is required.