Keep policy below the interface
Normalize MCP, API, CLI, SDK, and workflow ingress into one typed capability request so a new client does not create a weaker execution path.
For AI platform teams
One execution contract for every agent surface.
Onterix is being built as the governed execution layer beneath the agent clients, model providers, internal applications, and workflows your teams choose—so trust logic does not have to be rebuilt in each experience.
Agent clientsMCPCLI + SDKNative APIInternal agents
ONTERIX RUNTIMEOne evidence-bound transaction
IdentityData boundaryExact actionEvidence
Company systemsKnowledgeModelsWorkflows
Platform mandate
Enable useful agency without inheriting every control plane.
AI platform teams need a stable boundary between fast-changing agent experiences and the enterprise systems, credentials, policies, and operational responsibilities behind them.
Separate translation from authority
Adapters describe provider semantics. The runtime remains responsible for identity, data protection, policy, approval, credential resolution, execution, and evidence.
Treat enterprise work as durable
Represent waiting, approval, retry, uncertain outcomes, compensation, and reconciliation explicitly instead of reporting false synchronous success.
Change topology without changing meaning
The target contract keeps the same principal, capability, protection, approval, and evidence semantics across managed and customer-controlled modes.
Evidence-bound execution
Make the transaction the durable abstraction.
The target runtime carries one request from principal and capability resolution through destination-aware data protection, policy, exact-action approval, data-plane credential resolution, execution, and safe evidence.
01Normalize client intent into a typed capabilityingress
02Intersect actor, subject, client, and source authorityidentity
03Inspect source, model, tool, and delivery boundariesprotect
04Evaluate policy and bind any exact approvaldecide
05Resolve credentials in the active data planeexecute
06Return a result or durable run with evidenceobserve
Adoption path
Start with one blocked workflow—not a universal rollout.
A useful private-preview evaluation chooses a consequential action, identifies every system and control owner, and proves the boundary before expanding interfaces or connector breadth.
Define the real transaction
Name the actor, agent, client, sources, destination, payload, approval owner, expected result, and failure consequences.
Place the sensitive path
Decide which credentials, content, indexes, findings, workflow state, model routes, and evidence must remain customer-local.
Exercise change and failure
Test stale revisions, changed policy, transformed payloads, denied approval, provider failure, retries, and uncertain outcomes—not only the happy path.
Add clients after the boundary holds
Project the verified capability through additional interfaces without creating a parallel authorization or execution path.
Target customer-controlled plane
Keep content-bearing execution out of the shared control plane.
In strict customer-data-plane and fully self-hosted target modes, the assigned runtime is intended to own the sensitive execution path while the central service coordinates narrow metadata, signed configuration, and deployment control.
- Connector credentials and customer database secrets
- Source content, indexes, chunks, and embeddings
- Protection findings, token vaults, and quarantines
- Workflow payloads and execution-local state
- Customer model routes, KMS bindings, and audit sinks
- Direct routing to the assigned regional cell or customer runtime
STATUS
Onterix is in private-preview development. The common runtime, client compatibility, connector depth, regional placement, customer-hosted packaging, and end-to-end evidence claims described here remain subject to implementation and verification.
AI platform design partner
Bring the workflow your current agent stack cannot safely execute.
We will map its interfaces, systems, principals, data boundaries, approval semantics, deployment requirements, and proof obligations into one concrete target contract.