Risk-aware routing
Choose the least expensive approved model that meets task, quality, tool-use, privacy, region, and risk requirements.
Enterprise gateway product direction
One governed endpoint for enterprise AI.
Route each request to an approved, cost-effective model. Apply company policy before sensitive data leaves or tools execute. Retry safely without repeating side effects. Preserve evidence end to end.
agentaction.gatewaypolicy attached
MODEL
Classify → constrain → route
Approved provider · qualified model · budget · company context
ACTION
Authorize → execute once → assure
Exact payload · approval · durable state · provider evidence
01 / The enterprise gap
Inference control is not action authority.
Model gateways optimize cost and availability. IAM controls access. AgentAction adds the missing runtime decision for the exact tool call, payload, job state, approval, data boundary, and prior execution.
02 / Product promise
Route efficiently. Act deliberately.
The gateway packages the AgentAction boundary for enterprise adoption. It does not turn prompts into security policy or replace provider-side authorization.
Company policy
Apply provider, model, tool, resource, destination, data, budget, and versioned system-context controls from one governance plane.
Safe action replay
Bind idempotency to the exact request and return the prior provider result without repeating a refund, send, write, or deploy.
Action evidence
Link selection, authorization, approval, execution, replay, observation, and assessment without treating model output as authority.
03 / Company onboarding
Observe first. Enforce with evidence.
Start with one team, one workflow, and one consequential action class. Expand only after the policy and operating evidence are understood.
- 01
Connect
Bring an enterprise identity provider and existing model or tool credentials.
- 02
Observe
Discover models, tools, costs, data flows, repeated calls, and consequential actions without blocking traffic.
- 03
Apply policy
Start from a company baseline, then scope stricter rules to teams, workflows, models, tools, and destinations.
- 04
Enforce
Challenge or deny high-risk actions, route eligible work, and export correlated evidence.
04 / Deployment
Use the boundary where your traffic already flows.
Managed gateway
A hosted enterprise endpoint with tenant policy, durable state, approvals, and evidence.
Customer-controlled runtime
Run the enforcement data plane in a customer cloud or VPC while using a shared governance plane.
Existing gateway integration
Connect AgentAction as the decision and evidence service behind an MCP, API, or inference gateway.
05 / Proof, not promises
The boundary exists. The product surface is next.
Public code, fixtures, and tests remain the source of truth. Planned gateway capabilities stay labeled until an end-to-end demonstration exists.
- Hosted authorization, approvals, JIT grants, and tenant manifests
- PII and destination controls
- Idempotent provider-result replay
- Signed provider receipts and middleware
- MCP reference adapter and local guard mode
- Production gateway packaging and lifecycle
- Risk-aware inference routing and evaluation
- Managed, versioned company system context
- Customer-controlled deployment options
- Broader model, MCP, and agent-protocol compatibility
06 / Compatibility
Reuse the standards enterprises already trust.
AgentAction profiles established interfaces before proposing new vocabulary. Emerging work is tracked without presenting drafts as adopted standards.
Launch compatibility
- Model interfaces
- OpenAI Responses and Chat Completions; Anthropic Messages
- Agent tools
- MCP 2026-07-28 with versioned compatibility
- Identity and access
- OAuth, OpenID Connect, and MCP authorization
- Policy decisions
- OpenID AuthZEN Authorization API 1.0
- Telemetry
- W3C Trace Context and OpenTelemetry GenAI conventions
- Evidence
- JOSE JWS/JWKS, canonical request digests, and stable correlation
- Enforcement seams
- Envoy external authorization and agentgateway ExtMCP processors
Design compatibility
- A2A 1.0
- Agent-to-agent delegation and task governance
- WIMSE and SPIFFE
- Executing workload identity across trust domains
- ID-JAG and transaction tokens
- Downscoped human, agent, and transaction context
- Shared Signals
- Revocation, risk change, and kill-switch propagation
- Gateway API Inference Extension
- Customer-managed and self-hosted model routing
- SCITT and COSE Receipts
- Optional transparency for durable evidence
Compatibility describes an implementation target, not certification by OpenID, IETF, W3C, CNCF, the MCP project, A2A, or Kubernetes.
Start with one consequential workflow.