> ## Documentation Index
> Fetch the complete documentation index at: https://docs.holos.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Services and data residency

> What each Foundry service does, where it runs, and what (if anything) leaves your infrastructure.

A Foundry runner doesn't do one kind of work. Each Mac advertises a set of **services**, and the queue hands it only the job types it advertises and its credentials permit. This page is the reference for what those services are and, for each one, whether the material stays on your hardware.

<Note>
  Foundry is available to select customers only. [Contact us](mailto:hello@holos.io) to discuss dedicated runners and data-residency requirements.
</Note>

## The services

| Service                  | What it does                                                                                                                                        | Runs on            |
| ------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------ |
| **Object training**      | Trains the ARKit reference object that makes a 3D model [trackable](/dashboard/objects/tracking). Needs roughly 64 GB free disk and runs for hours. | Your Macs          |
| **Video prep**           | Derives a video capture's poster frame, scrubber previews, and speech audio track.                                                                  | Your Macs          |
| **Transcription**        | Turns a video capture's audio into a timecoded transcript.                                                                                          | Your Macs          |
| **Step extraction**      | Turns a capture's transcript into procedure steps.                                                                                                  | Your Macs          |
| **Pose extraction**      | Reconstructs a posed 3D body mesh from a picture capture.                                                                                           | Holos servers only |
| **Document conversion**  | Turns a reference PDF into a structured document its procedures can be extracted from.                                                              | Your Macs          |
| **Procedure extraction** | Turns a structured document into a draft procedure, steps and callouts included.                                                                    | Your Macs          |
| **Procedure alignment**  | Matches a capture against a procedure, so each recording either corroborates its steps or raises a proposal.                                        | Your Macs          |

Object training is the one service that *must* run on a Mac: it depends on the Apple toolchain, which is why Foundry is a fleet of Macs in the first place. Pose extraction is the opposite case: it's hosted inference, so it runs server-side and never on your hardware.

One service runs only on Holos servers for a different reason. The texture bake, which makes a large 3D scan light enough for a headset to render, needs a Blender build that exists for Linux but not for Apple silicon. It belongs to object processing rather than to a service you switch on, and no Mac is offered the work.

When your own runners are offline, Holos capacity picks up the capture services so uploads still finish processing.

## Choosing which services a runner takes

Each runner's credentials permit a set of services: a ceiling we set with you when your fleet is provisioned, and widen or narrow on request.

Within that ceiling, the Foundry app's **Services** switches (**Settings → Services**) let the person at the Mac decide what it actually takes on: a machine can train objects overnight and stay out of the capture pipeline, for instance. A service its credentials don't permit is marked in that window rather than hidden, so "why isn't this Mac taking transcription jobs" always has a visible answer.

<Note>
  Transcription requires macOS 26 or later, because it uses Apple's on-device speech models. On an older Mac the service is listed as unsupported and those jobs go elsewhere.
</Note>

## Where your data is processed

Some services do all their work on the machine that claimed the job. Others hand the material to an external provider, which is a decision you may need to account for. This distinction is a property of the **provider**, not the service: the same job type can be local or external depending on which provider ran it.

| Service              | Provider                               | Where it runs            | What leaves your infrastructure |
| -------------------- | -------------------------------------- | ------------------------ | ------------------------------- |
| Object training      | CreateML                               | Foundry runner           | Nothing                         |
| Video prep           | AVFoundation                           | Foundry runner           | Nothing                         |
| Transcription        | Apple Speech                           | Foundry runner           | Nothing                         |
| Transcription        | Whisper                                | [fal.ai](https://fal.ai) | Capture audio                   |
| Step extraction      | Apple on-device                        | Foundry runner           | Nothing                         |
| Step extraction      | Apple Private Cloud Compute            | Apple                    | Transcript text                 |
| Step extraction      | [Anthropic](https://www.anthropic.com) | Anthropic                | Transcript text                 |
| Pose extraction      | SAM 3D                                 | [fal.ai](https://fal.ai) | Capture photo and/or video      |
| Document conversion  | Docling                                | Foundry runner           | Nothing                         |
| Document conversion  | PDFKit                                 | Foundry runner           | Nothing                         |
| Procedure extraction | Apple on-device                        | Foundry runner           | Nothing                         |
| Procedure extraction | Apple Private Cloud Compute            | Apple                    | Document text, per section      |
| Procedure extraction | [Anthropic](https://www.anthropic.com) | Anthropic                | Document PDF and text           |
| Procedure alignment  | Apple on-device                        | Foundry runner           | Nothing                         |
| Procedure alignment  | Apple Private Cloud Compute            | Apple                    | Procedure and capture step text |
| Procedure alignment  | [Anthropic](https://www.anthropic.com) | Anthropic                | Procedure and capture step text |

The same table is in the dashboard, under **Where is my data processed?** on the [Runners](/foundry/runners) page, where it always reflects what your fleet currently runs.

The two rows naming [fal.ai](https://fal.ai) are served only by Holos-side capacity, never by your Macs. With your own runners online, transcription is handled by Apple's on-device models and the audio stays on the machine.

The three services that read language, step extraction, procedure extraction and alignment, each offer the same choice of provider, and the choice is what decides whether anything leaves. On Apple's on-device models the work happens on the Mac and nothing is sent. Apple Private Cloud Compute and Anthropic are both hosted, so the text goes to them wherever the job was claimed.

No provider credentials are ever stored on a runner, and runners only ever make outbound requests to Holos.

Every finished job records which provider did the work, so what left is answerable per job rather than in general. The dashboard labels each run in the [job queue](/foundry/jobs).

<Note>
  If your organization needs a service pinned to your own hardware, that's a fleet configuration we set up with you. [Talk to us](mailto:hello@holos.io) about residency requirements.
</Note>