Native Agent Runtime

    Your agent runtime.
    Not theirs.

    The agent runtime built into every DataInbox. Connect any DataInbox to any other, and to outside agents, without inheriting their dependencies, their cloud, or their failure modes.

    Your inbox is the runtime. External frameworks plug in as guests.

    Three guarantees, every deployment

    The runtime is the product. These three properties hold whether you run on SaaS, in your VPC, or fully air-gapped.

    Native, not bolted-on

    The runtime ships inside every DataInbox. No external orchestrator, no extra control plane, no vendor SDK in your critical path.

    Federated by design

    Any DataInbox can call any other DataInbox's agents over a typed contract. Multi-tenant, multi-region, multi-org. Verified Business Events on the wire.

    Sovereign by default

    Runs in your environment: EU, private cloud, self-hosted, air-gapped. External agents are guests, never owners. If a vendor goes down, your workflows keep running.

    Why native beats bolted-on

    Every external agent framework adds a runtime, a state store, a billing relationship, and a failure domain. A native runtime removes all four.

    External runtime (LangGraph, OpenAI Agents SDK, CrewAI)

    Runtime is part of the inbox. Nothing extra to deploy.

    Separate state store for agent memory

    Agent state lives next to the events, in the same governed store.

    Extra billing relationship with a framework vendor

    One product, one contract, one bill.

    New failure domain in your critical path

    One runtime to monitor. One audit log. One SLO.

    Federation, not centralization

    Any DataInbox can call any other DataInbox's agents. Each side keeps its own policies, residency, and audit log. The wire format is the Verified Business Event.

    DataInbox A
    EU, your VPC
    Verified Business Events
    Typed contract on the wire
    DataInbox B
    US, partner org
    OpenAI / Anthropic
    MCP servers
    Custom agents

    External agents attach as guests, on either side, through the same policy boundary.

    External agents as guests, never owners

    Plug in any external agent or framework through the policy boundary. They propose. The deterministic engine decides. Every call lands in the same audit log.

    OpenAI Agents

    Registered as guest tools through the policy boundary.

    Anthropic / Claude

    Called as a guest model. Proposals, never final decisions.

    MCP servers

    Any MCP-compatible tool plugs in over the standard protocol.

    LangGraph / CrewAI nodes

    Wrap a graph as a guest agent. Inputs and outputs validated.

    n8n / Zapier flows

    Trigger external automations as a side effect of an approved decision.

    Custom in-house agents

    Anything that speaks HTTP, gRPC, or MCP can be a guest.

    Run independently. Stay sovereign.

    No outbound dependency required to execute a rule. Lose a model provider, lose a framework vendor, lose your internet uplink, the inbox keeps running.

    • SaaS in the EU and US
    • Private cloud in your VPC
    • Self-hosted on-prem
    • Fully air-gapped deployments
    • Model providers are hot-swappable
    • Audit log stays where your data stays

    DataInbox Agent Runtime vs the alternatives

    Other frameworks assume you adopt their runtime. DataInbox inverts it.

    CapabilityDataInboxLangGraphOpenAI Agents SDKn8n agents
    Sovereign / self-hostablepartialpartial
    Federation between instances
    Deterministic policy enforcementpartial
    External agents as guestspartialpartial
    Hot-swap model providers
    Immutable audit log built in
    Runs without vendor cloudpartialpartial

    One audit log. Human, internal agent, external agent.

    Every call lands as an event with rule trace, approver, model used, and cost. Whether the actor was a person, your own agent, or a guest from OpenAI.

    Rule trace
    Which rule fired, which version
    Model trace
    Which provider, model, prompt version
    Decision trace
    Approved, rejected, escalated, by whom

    Frequently asked

    Do I need LangChain, LangGraph, or the OpenAI Agents SDK to run agents on DataInbox?

    No. The runtime is built into DataInbox. You can still use those frameworks if you want, by registering them as guest agents through the policy boundary, but they are never a hard dependency.

    What happens if OpenAI, Anthropic, or my model provider goes down?

    Your inbox keeps running. Rules, routing, audit, and federation are deterministic and local. Only the calls that target that specific provider fail. You can hot-swap to another model without code changes.

    Can two DataInboxes in different companies talk to each other?

    Yes. That is the federation model. Each side keeps its own policies, data residency, and audit log. Agents call each other through typed Verified Business Events, not raw APIs.

    Where does agent state live?

    Inside the DataInbox that owns the agent. Same store as the events, same backups, same residency, same audit. No external state service to manage.

    Is it really self-hostable, including air-gapped?

    Yes. SaaS in the EU and US, private cloud in your VPC, full self-hosted including air-gapped. The runtime has no outbound dependency required to execute a rule.

    Who makes the final decision, the LLM or the runtime?

    The deterministic policy engine. LLMs propose. The runtime decides, logs, and enforces. Every action is traceable to a rule and a model version.

    Run agents on your terms.

    One runtime. Federated by default. Sovereign by default. External agents welcome, never required.