A hybrid AI setup combines a work environment on your own system with a language model that runs locally or with an external provider. The agent works at your location, while the language model can use a provider such as OpenAI or Anthropic. So what exactly stays with you, and what leaves? This article goes through it component by component, including what goes to the provider with each task and how you set that boundary yourself.
These components run and are stored on a system at your location:
The Hermes environment itself, with its configuration, settings, and skills.
Your projects and files, such as code, documents, and a knowledge base.
Session history: earlier conversations, decisions, and context.
Access permissions: who can use the agent and with which rights.
Repositories can live where you choose, such as GitHub or an internal Git server. Which system holds the local part depends on your setup: a computer you already have or one you choose for it. Also read when you need a Linux computer for an AI agent.
The language model: the part that understands your instructions, makes a plan, and writes answers. That model can run with an external provider such as OpenAI or Anthropic, or locally on your own hardware. Hermes sends every message and every tool step through that main model. There are also smaller side tasks, such as summarizing a long conversation or looking at an image. By default, they use the same model, but you can pick a different model for each side task, including a local one. The documentation on configuring models explains how.
Short answer: each call only includes the context of the current task. Your files and earlier sessions stay local unless the agent opens or searches them for that task.
That context has three parts. First, the standing instructions for the agent: the project’s working rules, the short memory with fixed facts, and the list of available skills. Second, the conversation in the current session. Third, the output of the tools the agent used in that session, such as the text of a file it opened or the result of a search. A file the agent never opens is not sent. The documentation on prompt assembly shows how Hermes builds those instructions.
What the provider does with that context is set out in its own policy. OpenAI writes in its enterprise privacy commitments, which cover its API Platform among other products: “We do not train our models on your data by default.” According to the Anthropic Privacy Center, Anthropic does not use inputs or outputs from its commercial products, including the API, to train its models by default. The exception is when you send feedback or otherwise choose to allow it. Also check how long your provider keeps data; that differs by product and account.
| Aspect | Local | Cloud |
|---|---|---|
Hermes environment | Configuration, settings, skills | Does not run externally; the agent’s instructions are sent with each call |
Projects and files | Code, documents, knowledge base | Only what the agent reads for the task |
Session history | Earlier decisions and context | Not stored as the local session archive; sent context follows provider policies |
Access permissions | Who has access and with which permissions | Not applicable |
Language processing | Local model (if the hardware supports it) | External provider (OpenAI, Anthropic) |
You stay in control in three ways: you choose the model, you decide what the agent can open and use, and you work with your own provider account.
First, you choose the model. That can be a provider or a local model if your hardware can handle it. You can also switch per task: a sensitive task on the local model, the rest with the provider. For the hardware side, see what computer you need to run AI locally.
Next, you decide what the agent can open. During setup, you decide which folders, projects, and integrations the agent can access, and which tools are turned on. For an MCP integration, you also choose which of its tools Hermes loads. What the agent cannot open, it cannot send on.
Finally, you set up your own provider account. Usage and billing then run through you, and the provider’s policy applies directly to your account. Voltti has no access to your usage.
For a small-scale example, read about Sisu, our home agent with Hermes and Home Assistant: the device connections stay at home, while the language model uses an external service. A hybrid setup also helps when the cloud model stops. On October 22, 2026, the cloud model refused to continue in one of our own sessions, and a local model took over the work in the same session. Our article on a local model as a fallback explains how to set that up.
Read more about local AI for businesses or about security and privacy. Compare local and cloud AI in our article on local AI vs. cloud AI. Before you connect data, use our AI safety checklist for business data.
No. Each call only includes the context of the current task: the agent’s instructions, the conversation in that session, and what the agent read for that task. Your other files and earlier sessions stay local. What the provider does with that context is set out in its policy. Under that policy, OpenAI and Anthropic do not use data from their business APIs to train models by default.
Yes, with a local language model. Then no context goes to an external model provider. How much hardware that takes depends on the model; see what computer you need to run AI locally. A local model can be slower and less capable than a large cloud model. A middle ground is a local model as a fallback next to a cloud model. We advise you on this during the intake.
You do. Voltti advises you, but you set up your own model provider account. Usage and billing then run through you, and the provider’s policy applies directly to your account. You can see your own usage, and Voltti has no access to it.
