Is a private AI coding agent the same as a self-hosted model?
Not necessarily. You can keep the request path private by running a client against a managed model service in your own cloud account, with no model servers to operate.
Learn
What is a private AI coding agent?
A private AI coding agent reads a codebase, plans a change, edits files, runs checks and proposes the result, with the model call, the identity and the data staying inside infrastructure the customer controls.
Agent versus assistant
| AI coding assistant | AI coding agent | |
|---|---|---|
| Unit of work | A suggestion or an answer | A task: plan, change, verify, propose |
| Acts on the repository | You copy or accept the suggestion | Edits files and runs commands, within the permissions you grant |
| Needs | Good completions in the editor | Approvals, allowlists, isolation and a review step, because it acts |
What makes it private
“Private” is about who operates the path your code and prompts take. A private agent sends inference only to an endpoint you control, signs in with your own identity, keeps its data on the machine or in your own storage, and sends no telemetry to the vendor. Data sovereignty for AI coding covers which data is involved and where it should live.
Three ways to deploy one
| Model | What you run | Trade-off |
|---|---|---|
| Vendor SaaS | Nothing; the vendor operates the service | Fastest start; a vendor service sits in the request path |
| Self-hosted model | Model servers and GPUs, plus the agent | Full control; you size, buy and operate inference hardware |
| Client on a managed model service in your cloud | The agent as a client; the cloud provider serves the models in your account | No hardware to run; you rely on your cloud provider’s model service |
Topsoil takes the third route on Amazon Bedrock: a client on the machine that calls Bedrock in your own AWS account, commercial or AWS GovCloud (US). See how Topsoil deploys.
Questions to ask any vendor
- Which endpoints does the client talk to, and can I restrict them?
- Is there a vendor-operated service, prompt store or telemetry stream?
- Where are approvals enforced: in the model’s instructions, or in the client where actions happen?
- Does the agent merge its own work, or propose it for review?
- What stops spend, and what is only an estimate?
Questions
Frequently asked questions.
Does a private agent mean no internet access?
No. Most private agents still need to reach a model endpoint. See the explainer on disconnected (air-gapped) networks for what that means.
See it against your own codebase.
A pilot runs in your own AWS account, with your models and your material, and ends with findings you can act on.