Plan-only by default
Every Terraform output ships with prevent_destroy and an apply guard. The default execution mode is dry-run.
Security
Fabric Projects is built so that the failure mode of a poorly-prompted agent is a closed PR, not a misconfigured cloud router. Our security model leans on Equinix's existing identity and approval surfaces — we just make them legible to the agent before it asks.
Every Terraform output ships with prevent_destroy and an apply guard. The default execution mode is dry-run.
Agents authenticate as network-agent-mcp@ users with project, metro, and resource scopes. Personal admin identities are never used for OAuth consent.
Five blocking gates run before any mutating tool: ports/devices, observability permission, developer app, capacity confirmation, provider quote.
The agent reads commercial estimates from a versioned pricing.md contract. Anything missing is reported as quote_required rather than fabricated.
Every plan run captures the prompt, the read tools, the blocked mutating tools, the timestamp, and the dedicated MCP user. Logs export as JSON or to a customer SIEM webhook.
Output is shaped to be reviewable in a PR — modules, lifecycle blocks, apply guards, and explicit quote-status fields the reviewer can grep for before merging.
The local workbench keeps everything on disk. The connected agent runs against your Equinix tenant with a read-only MCP scope by default. Nothing goes to a third-party inference provider unless you explicitly route through Equinix Fabric — at which point the path is private and observable.
in audit · Q3 2026
scoped · 2027
BAA on request
via Equinix Government Cloud
none for the local demo
Security review used to take weeks because everyone wanted to see what the network change would actually do. The plan-only output and audit log let us close that loop in a single afternoon.