Trust & data

Inspect execution. Control your data.

Inspect model and tool calls, keep each user’s data separate, and export or delete your data through the API. Agent state is hosted in Europe; the details below explain the boundaries.

Four properties

What the record gives you.

What the trace, the scoping, the residency, and the export path each do, and where each one stops.

Every action on the record

Each run is a persisted, replayable sequence of steps: every model call, tool invocation, approval, and euro. Open one later and see exactly what happened, and what it cost.

run.started → tool.executing → approval.required → run.completed

Separate data for each user

Each user’s assistant has separate memory, tool connections, and history. User-scoped tokens restrict access to that assistant’s data; project administration uses a separate token.

scope: smith · per-project isolation

Agent state stays in Europe

Smith memory, threads, runs, and traces are stored and processed on EU infrastructure, because that's where the platform runs. Model calls go wherever the model is served: bring your own key and you pick the provider and region yourself. The subprocessor list names each one.

agent state hosted in the EU

Your data, yours to take

Read it through the same API you write it with, export it, or delete it. There's no proprietary store you can't get back out of.

GET /v1/... · export · delete
No lock-in

Infrastructure you can leave.

You can take your data out and run your integration elsewhere. Here’s what that costs you.

  • Standards in

    OpenAI-compatible API, MCP for tools.

  • Your data out

    Full export through the same API, anytime.

  • Partial exit, not zero

    The model surface is standard, so inference moves as-is. Isolation, memory, approvals, and the trace are ours; those you'd rebuild.

What we’re upfront about

We document what ships.

Three things worth knowing before you commit.

  • Webhooks are one signed attempt today

    We deliver each lifecycle event once, signed for you to verify. There's no automatic retry yet, so the durable source of truth is the append-only events feed rather than the webhook. Poll it to reconcile.

  • No SOC 2, no ISO 27001

    We hold neither certification and have no audit date to give you. What exists today is the architecture above, the subprocessor list, and the DPA.

  • Model calls can leave the EU

    Agent state is stored and processed in the EU. Managed model access routes to whichever region serves the model, which for some models is the United States. Bring your own key and you choose; either way the subprocessor list names the processor and its location.

Read the events & webhooks docs

Connect your application with scoped access.

Configure project credentials, user-level access, and approval rules. Inspect model and tool execution through the console and API.