Status: implemented
Version: latest
Review: source-backed
Last scanned: 2026-06-25T00:00:00Z
Review required: false
Vendor and integration reference
How Hermes, NVIDIA, Stripe, Gmail, Jido, HubSpot, Fly.io, Tigris, and fallback providers map into Trinity.
Integration thesis
Trinity is not a wrapper around one model or one payment API. It is a control plane that lets an agent operate a revenue workflow while every external system remains behind account scope, credential readiness, policy checks, human approval, idempotency, and audit proof.
The vendor integrations are deliberately separated by responsibility:
- Hermes provides the hosted agent runtime and skill/memory posture.
- NVIDIA provides model intelligence, Nemotron routes, safety review, and official endpoint proof.
- Stripe provides revenue mode, checkout, webhook proof, and future guarded spend/provisioning.
- Gmail provides account mailbox OAuth, project alias selection, draft/send/reply rails, thread inspection, labels, and ingestion after OAuth, credential-vault, policy, and approval gates pass.
- Jido provides a narrow typed policy seam for content-firewall execution.
- HubSpot is optional sync; the built-in CRM remains the source of truth.
- Fly.io, Postgres, Oban, and Tigris support deployment, durability, jobs, and artifact storage.
Vendor map
| Vendor or system | Trinity role | Implementation boundary | Proof records | Official references |
|---|---|---|---|---|
| Nous Hermes Agent | Hosted agent runtime for planning, drafting, action intents, and skills |
HermesRunWorker, HermesRuntimeClient, hermes_agent_profiles, scoped hermes_skills
| Hermes session/message, run event, skill link, profile snapshot | Hermes docs , profiles , skills |
| NVIDIA NIM and Nemotron | Model scoring, copy safety, classification, QA, and high-value reasoning |
ModelRouter, Tools.Nemotron, DraftSafetyReview, DeliverableQA
|
ModelRoute, ModelInvocation, ModelDecision, run event, audit event
| NIM overview , NIM API |
| NVIDIA NeMo and NemoClaw posture | Guardrail and sandbox posture for model/tool safety |
NemoClaw, RuntimePosturePolicy, ContentFirewallPolicy
| Safety setting, policy denial, sandbox execution proof | NeMo Guardrails , rail types |
| Stripe | Revenue mode, checkout, signed webhook processing, and guarded agent spend |
Revenue, StripeWebhookController, Tools.Stripe, StripeAutonomousSkills
|
CheckoutSession, StripeWebhookEvent, Order, RevenueEvent, Deal, ToolCall
| Checkout Sessions , webhooks , API reference |
| Gmail and Google OAuth | Account mailbox OAuth, project alias selection, draft creation, approved sends, thread/list/label calls, ingestion, and replies |
GoogleOAuthController, TokenStore, GmailClient, GmailAdapter, ProjectGmailSetting, mailbox and alias records
| Gmail mailbox, project Gmail setting, sending alias, draft, thread/message, ToolCall, ApprovalEvent | Gmail scopes , drafts , drafts.create , messages.send |
| Jido | Typed action seam for content-firewall payload checks |
ContentFirewall.PayloadCheckAction, JidoRuntime, ContentFirewallPolicy
| Firewall decision metadata, policy denial, audit/run events | Jido docs , Jido Action , concepts |
| HubSpot | Optional CRM export/sync when approved |
HubSpot.Sync, HubSpotObjectMapping, ToolRouter adapter
| Mapping rows, ToolCall, audit event, CRM record refs | HubSpot private app and CRM docs from the connected workspace |
| OpenRouter and local NIM | Explicit fallback-only model routes |
Disabled fallback ModelRoute rows
|
Fallback-labeled ModelDecision, never relabeled as official NVIDIA proof
| OpenRouter docs |
| Fly.io | Production hosting and release flow |
Dockerfile, fly.toml, release commands, /healthz, /readyz
| Deployment logs, health checks, migration result | Fly Elixir docs , Phoenix Fly guide |
| Tigris or S3-compatible storage | Private documents, attachments, and generated artifacts when configured | Storage config, document/artifact controllers, signed download paths | Project document, artifact, object key, release status | Tigris docs , S3 compatibility |
Sponsor-facing implementation details
| Sponsor layer | What Trinity showcases | Why it matters |
|---|---|---|
| Nous Hermes | Project/account-scoped Hermes profiles and skills prevent shared agent memory across tenants | The agent can learn and reuse work without making global skill leakage the default |
| NVIDIA | Official NIM/Nemotron routes, safety route, showcase route, and durable ModelDecision evidence | Model work is visible, validated, and fail-closed rather than an opaque text response |
| Stripe | A real $1 snapshot product can become paid revenue proof and later guarded spend/provisioning | Agents can earn, request purchases, and fulfill operations under human control |
| Google/Gmail | Drafts and sends are separate stages with account OAuth, project alias binding, vault-backed tokens, safety review, and approval | Outreach automation is useful without letting an agent send unchecked email or another project's mailbox |
| Jido | Jido is used where it improves quality: typed content-firewall execution around poisoned inputs | The system adds agent framework value without replacing durable Phoenix/Postgres authority |
Secret and readiness boundaries
Trinity stores safe provider metadata and readiness posture, not raw credentials. Required provider slots are surfaced as setup status, while values remain in environment configuration or credential-vault storage paths.
| Boundary | What is visible | What is never public |
|---|---|---|
| Provider catalog | Provider names, setup status, credential slot labels, official/fallback status | API keys, OAuth refresh tokens, webhook secrets |
| Hermes profile | Profile kind, model provider, base profile, required slot metadata | Runtime key, raw project secret values, private model context |
| Model proof | Route name, provider, schema version, status, hashes, summary | Raw prompts, hidden reasoning, credentials, private payloads |
| Tool proof | Provider, operation, idempotency key, policy status, redacted request/response | Bearer tokens, customer secrets, unreleased inbound content |
Related pages
Official references
| System | Use in Trinity | Official docs |
|---|---|---|
| Hermes | Hosted runtime, profiles, and skills | Hermes Agent docs |
| NVIDIA | NIM/Nemotron route and safety proof | NIM LLM API reference |
| Stripe | Revenue and future spend rails | Stripe webhooks |
| Gmail | Drafts, sends, aliases, and replies | Gmail drafts guide |
| Jido | Content-firewall action seam | Jido docs |
| Fly.io | Elixir production deployment | Fly Elixir docs |
| Tigris | S3-compatible artifact storage | Tigris docs |
Source paths
lib/autonomous_agency/integrations/provider_catalog.exlib/autonomous_agency/hermes/agent_profile.exlib/autonomous_agency/hermes/skill_registry.exlib/autonomous_agency/ai/model_router.exlib/autonomous_agency/tools/tool_router.exlib/autonomous_agency/revenue.exlib/autonomous_agency/tools/gmail_adapter.exlib/autonomous_agency/email/gmail_client.exlib/autonomous_agency/google/token_store.exlib/autonomous_agency/email/project_gmail_setting.exlib/autonomous_agency/security/content_firewall.exdocs/plans/2026-06-29-current-production-readiness-plan.md