AI Gateway Governance
AI gateway governance applies policy before model, template, or tool traffic leaves Eden. It is used for sensitive document workflows, internal research, support automation, operations agents, and other regulated AI workflows.
This is an Adam feature. Eve can govern model providers as endpoints; Adam governs the AI experience where licensed users, agents, or gateway API keys use models and tools against Eve endpoints. Gateway API keys are the Adam control surface for model-use policy, route selection, budgets, tool calling, and AI evidence; seat and API-key entitlements are governed by the applicable Eden license.
The gateway accepts OpenAI-compatible traffic and applies Eden policy as a pre-egress layer. Applications keep using a familiar model API shape while Adam owns the LLM interface and tool-calling behavior, and Eve enforces the endpoint access, template access, tool access, credentials, shared PII or masking controls, and telemetry underneath.
Control Stack
| Layer | Role |
|---|---|
| Gateway API key policy | Stores per-seat or per-agent route, budget, model, PII, lookup, and metadata policy for OpenAI-compatible traffic. |
| PII controls | Combine built-in detection, organization dictionaries, key dictionaries, and request-scoped lookup terms. |
| Template-backed lookup | Runs saved Read templates in sensitive mode to derive request-scoped terms from trusted records. |
| Content policies | Detect restricted topics, secrets, payment instructions, projects, watchlists, or other configured terms. |
| Policy bindings | Attach profiles to organizations, users, agents, auth groups, gateway keys, endpoints, routes, and work metadata. |
| Access boundaries | Check subject/resource combinations before model, template, or tool egress. |
| Promptless evidence | Stores simulations, approvals, exports, lineage, replay jobs, and feature records without prompt text. |
Runtime Flow
- A client sends OpenAI-compatible chat or responses traffic to Eden.
- Eden authenticates the gateway key and loads the persisted policy.
- PII, content policy, and access-boundary controls evaluate before provider egress.
- Template-backed lookups run only through permission-checked Eden
Readtemplates. - Eden forwards only the allowed and masked upstream request.
- Telemetry records safe identifiers, statuses, counts, route data, cost data, and work metadata.
Configuration Surfaces
| Surface | Purpose |
|---|---|
| Gateway API keys | Persist model routes, budgets, PII dictionaries, lookup templates, metadata policy, and provider limits. |
| LLM endpoints | Store provider credentials and connection settings for OpenAI, Anthropic, Ollama, OpenRouter, Azure OpenAI, and compatible providers. |
| Policy profiles | Define reusable content and risk categories. |
| Policy bindings | Attach profiles to subjects, routes, keys, endpoints, and work metadata. |
| Access boundaries | Decide which subjects can send traffic toward resources. |
| Promptless evidence | Store safe policy decisions, approvals, exports, and replay artifacts. |
Composition
Controls are additive. A single request can be routed by model policy, enriched by template lookup, redacted by PII governance, checked against content profiles, blocked by an access boundary, and recorded in promptless evidence. Eden evaluates the controls once before egress so the upstream provider only sees the final allowed and masked request.
Rollout Guidance
Start by configuring route and budget policy, then add PII governance and content profiles in audit-oriented mode. Use promptless simulations and safe evidence records to review decisions before enabling blocking for production traffic.
Agent Runtime Integrations
Agent runtimes are clients of the Adam gateway and use separately governed Eve endpoints. Their configuration determines the native request protocol and the attribution Eden can record. Eden documents all supported agent integrations, including Codex, Claude Code, Hermes, and compatible agent clients.
Use a distinct gateway key for each independently operated agent identity when policies, budgets, routes, or evidence need to be separated. Application integrations such as GitHub, Google Workspace, and Salesforce remain Eve endpoints; they may be called by agents, workflows, APIs, or users under the same endpoint authorization model.
Release Guarantees
- Existing dictionary and regex PII behavior remains available when no lookup templates are configured.
- Lookup-derived terms are request-scoped and are not persisted into dictionaries.
- Sensitive template execution suppresses rendered request and response payload logging.
- Runtime stale config, revoked access, missing required inputs, and malformed persisted policy fail closed.
- Telemetry never includes raw prompt text, lookup input values, lookup output values, or matched PII values.