Does Thazen host anything in the request path?
No. Topsoil is a client. Inference goes to the Bedrock, custom HTTPS or gateway endpoint you configure, and there is no Thazen-operated service, prompt store or telemetry stream.
Deployment
A private AI coding agent that runs in your own AWS account.
Topsoil is a client. It calls the Amazon Bedrock, custom HTTPS or gateway endpoint you configure, signs in with the identity you already use, and keeps its data on the machine or in your own S3 bucket. There is no Topsoil service to deploy, and none in the request path.
How the data flows
Every arrow in this picture ends in infrastructure you operate. Thazen is not on it.
Your machineTopsoilTerminal, VS Code bridge or desktop app. Chats, credentials and tool output stay here.
Your network pathPublic, custom HTTPS or PrivateLinkThe endpoint you set in Settings, signed with your identity.
Your AWS accountAmazon BedrockCommercial regions or AWS GovCloud (US). Plus your S3 bucket for the context library and budget policy.
There is no Thazen-operated inference service, prompt store, analytics or telemetry stream, and no crash reporter or font or CDN fetch.
What leaves the machine, and what never does
| Data | Where it goes |
|---|---|
| Prompts and the context you attach | To the Bedrock, custom HTTPS or gateway endpoint you configured, and nowhere else. |
| Project files in the context library | Your own S3 bucket, under a per-user prefix that IAM enforces. Users cannot read or list each other’s files. |
| On-device search index | Built locally with Amazon Titan Text Embeddings v2 in your account. It honours .gitignore and is not uploaded anywhere else. |
| Standalone chats, credentials, terminal sessions, raw tool output | Stay on the machine that runs them. Desktop credentials are held in the operating system’s keychain. |
| Logs | Anything that can hold a key, token or session is redacted before it is written. If your account turns on Bedrock invocation logging, those logs are yours, in your account. |
| Updates | Where an update channel is configured, it is your own bucket, not ours. |
That is what data sovereignty looks like in practice: your account, your Region, your keys. Data sovereignty for AI coding, explained.
Identity: who is calling
Topsoil signs requests with the AWS identity you already manage, so access, logging and retention stay under the policies you already run.
- IAM roles and access keys, or named AWS profiles (role and
credential_processprofiles are supported). - IAM Identity Center (SSO) through
aws sso login. - Bedrock API keys, which an organisation can deny with a service control policy. Signing in with IAM does not need them.
- A bearer token for a customer-hosted gateway, for the browser app.
The IAM policies for commercial and AWS GovCloud (US) partitions ship with Topsoil.
Network paths
| Where the client runs | Endpoint | Notes |
|---|---|---|
| A laptop on the internet | Public regional Bedrock runtime endpoint | The default. |
| On a VPN or Direct Connect that resolves VPC DNS | Public, or the private DNS name | Only where the machine has a route to the interface endpoint. |
| A workstation inside the VPC | The interface endpoint (AWS PrivateLink) URL | Settings warns that this hostname only works inside the VPC. |
| The browser app | Your own gateway origin | Bearer token through a customer-controlled gateway. |
AWS documents PrivateLink for Bedrock, including a FIPS runtime endpoint, in its user guide.
No GPUs to run
Self-hosting a model means sizing, buying and operating inference hardware, then keeping the model current. Topsoil takes a different route: it uses the models Amazon Bedrock already serves in your account, so there is no model server, cluster or GPU fleet to stand up, and model choice grows as Bedrock’s does. More on building on Bedrock.
Setting up the account
One CloudFormation stack, deployed once per account and region, creates the context-library bucket, the user policy as a managed policy, an unattached Bedrock deny policy and, when you give it an amount and an address, a monthly AWS budget wired to it. Three steps stay manual because AWS does not let a template do them: enabling Bedrock models, assigning Identity Center permission sets, and deciding which roles get the policy.
Spend limits
Monthly user and account budgets are checked before each model call, so a request over the limit is not sent. A budget Topsoil enforces stops Topsoil, not Bedrock, so an AWS budget action your admin deploys is the backstop. How the three layers work.
Where it runs
A terminal client for macOS, Windows and Linux; a VS Code bridge that lends the terminal agent your editor’s context; a desktop app for macOS and Windows; and an optional browser app you host yourself. See the tour.
Questions
Frequently asked questions.
Can Topsoil use AWS PrivateLink?
Yes. The desktop app supports public Bedrock endpoints, custom HTTPS endpoints and reachable PrivateLink and VPC paths. The signing region must match the endpoint.
Do we need GPUs or a model server?
No. Topsoil uses the models Amazon Bedrock serves in your account, so there is no inference hardware to run.
What do we deploy in AWS?
One CloudFormation stack per account and region for the context library, the user policy and the optional budget backstop. You then enable your Bedrock models and attach the policy where you want it.
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.