Third-party integrations — one vendor account per environment

Neon Law Navigator talks to a handful of external services. They fall into two kinds:

• Binding vendors perform real, billable, or legally binding actions on the firm's behalf — DocuSign for e-signature, Xero for accounting and billing. For every such vendor we keep two separate vendor apps/accounts: a development (sandbox) account used for local dev and CI, and a production account used only in prod. This mirrors each vendor's own recommended setup — DocuSign's demo vs. production environments, Xero's demo company vs. live organisation — and it is the default way we develop against any binding third party.
• Platform services are the cloud infrastructure the app runs on — durable execution, object storage, the database, identity, the agent-router LLM, and outbound/inbound email. These don't take legally binding actions, so they don't need the two-account split; a fork points them at its own project / account through the same <VENDOR>_* env variables.

The full catalog below lists every external service the application code itself dials. Purely operational layers that sit above the env-var interface — Doppler (secret values), DNSimple (DNS) — are deliberately out of scope here: they are not code dependencies, and a fork can swap them freely.

Why two accounts

• No legal or financial weight in dev. A test envelope or a draft invoice created against the sandbox account is not a binding signature or a real ledger entry. A leaked dev key cannot mint a production signature request.
• Clean books and clean signers. Test data stays out of the real accounting ledger and off real signers' inboxes. Self-testable forks. An OSS adopter can stand up their own sandbox account and exercise the full flow without touching a real account or paying for live API calls.

Deployment selector and credential separation

The narrow deployment selector is exact: NAVIGATOR_ENVIRONMENT=staging selects staging, while empty or unset selects production. Mixed case, whitespace, and every other nonempty value fail configuration parsing. It controls infrastructure safety checks only.

• .env holds the sandbox credentials and is auto-loaded on startup, so local dev and cargo test run against the vendor sandbox by default.

• .env.production holds the production credentials. It is gitignored by the .env.* rule and never committed. To run against production locally, source it over the defaults before launching the binary:

  set -a; source .env.production; set +a
  

Both sources use the same application variable names (DOCUSIGN_*, XERO_*, …). The operator also sets NAVIGATOR_CREDENTIAL_ENVIRONMENT to staging or production; startup rejects a mismatch. Cloud staging receives the non-production set from a staging-labeled Kubernetes Secret, while production receives a distinct production Secret.

Normal staging requires real non-production SendGrid and DocuSign demo configuration. Only the explicit NAVIGATOR_CI_HARNESS=1 staging test surface may use in-process fakes; production rejects that flag.

Current integrations

Notes:

• Xero ↔ Mercury. Xero reconciles against the firm's bank (Mercury) inside Xero itself. Neon Law Navigator never speaks to Mercury — our only integration boundary is the Xero API.
• Google Cloud is several spec-compliant touchpoints, not one SDK. Object storage goes through the cloud crate's StorageService trait (GCS in prod, Garage in dev); the database is vanilla Postgres over DATABASE_URL (Cloud SQL in prod); OIDC is Google Identity validated against GOOGLE_OAUTH_*; the per-Project archive is Drive REST v3. See cloud/README.md for the full resource map.
• Vertex AI is pluggable. The router is the web::agent_router::AgentRouter trait — GeminiRouter (Vertex AI) in prod, NullRouter in KIND. Swapping to another LLM means a new impl, not a new vendor account.

When you add a binding vendor, follow the two-account shape: create the sandbox + production accounts, add a <VENDOR>_* block to .env.example that references this convention, put sandbox credentials in .env and production credentials in .env.production, and fall back to a stub when the vendor is unconfigured. A platform service needs only its own <VENDOR>_* block and a stub/local equivalent (Garage, in-cluster Postgres, the NullRouter) so a fresh checkout boots and self-tests without any cloud account.

Not in this catalog — and why

A few external-looking things are deliberately absent. They are not third-party SaaS vendors, so the per-environment account convention does not apply to them:

• OPA (Open Policy Agent) is first-party infrastructure you self-host, not a vendor. The same OPA container runs in both environments — a sidecar in the web pod in prod (examples/deploy/k8s/gke/patches/web-resources.yaml) and the in-cluster service in KIND (k8s/base/opa/opa.yaml), reached via NAVIGATOR_OPA_URL. There is no "OPA account" to sign up for.
• OIDC identity is already the Google Cloud row. In production, sign-in is Google Identity (validated against GOOGLE_OAUTH_*) — counted under Google Cloud above, not as a separate vendor. Keycloak is only its local stand-in (the dev/KIND OIDC provider), exactly as Garage stands in for GCS and in-cluster Postgres for Cloud SQL. The identity provider is pluggable and spec-compliant either way.
• Doppler and DNSimple sit above the env-var interface — Doppler holds secret values (the app reads plain env vars; see secrets-doppler.md), DNSimple holds DNS records. Neither is a code dependency, and a fork can swap both freely.

Related

• .env.example — the canonical per-variable reference; this convention is stated in its top "Conventions" block. oss-install.md — the install walkthrough's env-configuration step. env-driven-devx.md — the broader "one config surface, three audiences" env philosophy.