01
Model flexibility
Use the Amazon Bedrock models your account enables. The list differs by region, including AWS GovCloud (US). A model becomes available the day you turn it on in your account.
Private AI on Amazon Bedrock · For defense and regulated teams
The private AI platform for engineering,
in your own AWS account
Topsoil helps teams design, build and review software, cloud infrastructure as code and CI changes with AI that runs inside your own AWS account on Amazon Bedrock.
It runs in a commercial region or AWS GovCloud (US), with no GPUs to run and no vendor in the request path. Your repositories, plans and documents stay in infrastructure you control.
Inference runs in your account. Topsoil keeps no prompts and collects no telemetry.
Topsoil on Grok 4.7, fixing a real bug. It reproduces the failure with a new test, then shows the exact fix and waits for your keystroke. The work happens in its own git worktree, with the running cost in the footer.
- DesignWeigh the options and settle on an approach before the plan names a file.
- SoftwareTake an issue to a pull request in an isolated git worktree, with a test that has to fail first.
- InfrastructurePlan and edit Terraform, CloudFormation and pipeline files in git under the same gates.
- ReviewA second model checks the change, and a person always makes the merge.
02
An AI platform for engineering in your own AWS account
Topsoil runs against your account, your region, your network path and your IAM policy. Software, infrastructure as code and CI configuration stay under the same plan-first gates, because the work never passes through a platform we operate.
03
No external vendor telemetry
There is no analytics pipeline, no usage beacon and no prompt collection leaving your environment. Nothing about your code, your documents or your people is reported anywhere.
Topsoil runs no vendor-operated control plane in the path, so your own account and IAM policy decide where your work goes.
Data sovereignty
Your code stays in your AWS account
Topsoil runs in your own AWS account and the Region you choose, on Amazon Bedrock. Your code, prompts and the model’s outputs travel only between your machine and the Bedrock endpoint you configure, so they stay under your accounts, your IAM policy and your retention rules.
Your machineCode and outputsSource, chats, credentials and tool output stay on the machine that runs Topsoil.
Your RegionCommercial or AWS GovCloud (US)You pick the Region and the model. Requests are signed with your identity.
Your AWS accountAmazon BedrockServes the model in your account. The context library lives in your own S3 bucket.
No Topsoil service, prompt store or telemetry stream sits anywhere in this path. What data sovereignty means for AI coding
Where it runs
One agent in the terminal, VS Code and the desktop app
Engineers live in a terminal and an editor, so Topsoil starts there. Analysts, marketers and proposal teams live in an application window. Topsoil is the same product and the same governance in all of them.
Terminal
The TUI
The whole agent in a terminal, on a laptop, a build box, a bastion host or a locked-down workstation. Same models, same permission model, no desktop required.
Editor
VS Code extension
A bridge to the terminal agent, with no second agent. It lends the terminal agent your unsaved buffers, selection, open files and diagnostics, then opens a diff of what the turn wrote. The agent keeps running in the integrated terminal.
Desktop
Native application
A full workspace for macOS and Windows: projects, a context library, documents, diagrams, shared channels and a built-in Terminal. The natural home for people who do not live in a terminal.
The same agent, in VS Code. Topsoil runs in the integrated terminal on its own git worktree; when the turn ends, the extension opens a side-by-side diff of every file it wrote. Your main checkout is untouched until you merge.
Inside the workspace
You can always see what the model was given
Every answer shows its sources. The side panel lists the working folder, each attached file, how much of the context window is spent and what the turn cost. You can check the context at any point.
- The working folder, file by file, with nothing implicit.
- A token budget you can watch as you go, before it shows up in a bill.
- Per-turn cost, shown as the turn happens.
Working folder9 files
~/Projects/route-planner
- src/eta/cache.ts
- src/eta/reroute.ts
- src/routes/types.ts
- src/server.ts
- test/eta-cache.test.ts
3 project files · ~462 / 50k tokens1% of budget
this turn2,975 in · 251 out · ~$0.013
The same panel in the desktop app. The answer cites the line it read and waits for you to approve the fix. The footer shows what the turn cost.
Software and AI engineering
An agent that makes the change and checks it
Topsoil reads the repository, proposes the change, runs the command, opens the diff and checks the result. Anything that touches your machine waits behind a permission gate.
Codebase
Real workspace context
Inspect and edit files, follow language-server intelligence, and search a local semantic
index that respects .gitignore and never uploads your source to build itself.
Execution
Approved shell and Git
Run commands and Git operations under allowlists and path protections. Commands run without a shell, exactly as shown to you before they run.
Workflow
Multi-model stages
Route explore, design, plan, implement, verify, review and document to different models. Use a fast model to search and a strong one to reason, within a single piece of work. Explore and design run by default and can be switched off.
Delivery
Backlog to pull request
/task takes an issue, /goal works until a condition holds and
/loop reruns on an interval. Each shows you the plan, works in an isolated
worktree and proposes a change for a person to review. Topsoil does not approve or merge.
Backlogs can be GitHub Issues, self-managed GitLab, or Jira Data Center with Bitbucket Data
Center inside your own network.
Verification
Proof that the fix works
Verify runs the changed test with the rest of the work reverted and requires it to fail
first. A test that passes without the change does not count as done. After an ordinary turn
that writes a test, Topsoil makes the same check and tells you. It covers JavaScript and
TypeScript (Vitest, Jest, node --test), Go, Python (pytest), Rust integration
tests, and C and C++ built with CMake and CTest. A project built another way declares its
commands in .topsoil/tests.json; tests in other frameworks are not checked.
Integration
MCP and your systems
Add MCP servers over stdio, HTTP or WebSocket, enable them per project, and keep every secret host-side where the model never sees it.
Limits
A ceiling on every run
A run budget bounds each pass, and a loop ends itself after three fires that propose nothing or when its lifetime budget is spent. An unattended run will not start without a ceiling.
Scheduling
Exit codes a scheduler can read
A failed run exits non-zero, and a budget stop has a code of its own, so cron, CI and other schedulers can tell “raise the budget” from “the tests failed”.
Errors
Failures that say what to do
In the terminal a model error prints what happened and then the fix: a missing model subscription, an expired sign-in, a step that has to be done from the commercial AWS account, an untrusted certificate.
Done means proven. The suite passes, and Topsoil checks in a scratch copy that the new test fails without the change.
Control
You approve every edit and command
Plan mode stays read-only until you approve. You see every edit and command first, with the diff, the path and the exact command. Then you decide whether that approval covers one path, one tool or the rest of the session.
Human-in-the-loop approvals for every edit and command
The Topsoil client enforces these gates at the point a file is written or a command runs. The model is never asked to police itself, so text planted in an issue or a web page cannot talk its way past them. See where each control lives.
- Plan-first by default. Read-only until explicitly released.
- Tool risk classes and per-path protections you set per project.
- Command allowlists, with commands run without a shell.
- Agent features are off until someone turns them on deliberately.
- Spend limits. Monthly user and account budgets are checked before each model call.
src/eta/cache.ts (+1 −1)
− const entry = etaCache.get(route.id);+ const entry = etaCache.get(`${route.id}:${route.revision}`);
Accept once1
Rejectesc
Always this path
Always this tool3
The approval card. The exact lines wait for you. y allows them once, and a allows the tool for the rest of the session.
$ topsoil models Models in us-gov-west-1 (example) Meta us-gov.meta.llama3-3-70b-instruct-v1:0 tools xAI us-gov.xai.grok-4.6 tools · images us-gov.xai.grok-4.7 tools · images * xai.grok-4.7 tools · images * default model
Model flexibility
Any model your Bedrock account enables
Topsoil lists the models your account can call, groups them by vendor, and lets a person pick from that list. Change the default for a session, a project or a workflow stage. The list differs by region.
- Your region decides the catalog. AWS publishes which models are available, including in AWS GovCloud (US).
- Your account decides. Enable a model in your account and it appears.
- Different models per stage, so cost and capability match the task.
- No lock-in to a single provider's roadmap, pricing or availability.
Why not self-host a model?
Self-hosting means sizing, buying and operating GPUs, then keeping the models current. Topsoil builds on the models Amazon Bedrock already serves in your account instead, so there is no inference hardware to run and your model choice grows as Bedrock's does. More on building on Bedrock.
Privacy, security and data sovereignty
AI for defense, AWS GovCloud (US) and regulated engineering teams
Topsoil was built for organizations that cannot send engineering context or commercial documents to a multi-tenant platform: defense and federal programs, regulated industries and anyone whose data has a jurisdiction.
Residency
The data never leaves
Prompts, source, documents and output stay inside the account, region and network path you configure. There is no Topsoil-operated service in the request path and no prompt store anywhere outside your environment.
Telemetry
Nothing is reported
Topsoil sends no analytics, usage beacons or crash reports with your content. Topsoil was built without a telemetry pipeline, so there is nothing to switch off.
Identity
The controls you already run
Authenticate with IAM roles and keys, AWS profiles, IAM Identity Center (SSO), or a gateway bearer token. Access, logging and retention stay under your existing policy.
Isolation
Private network paths
Public endpoints, custom HTTPS endpoints, or reachable PrivateLink and VPC paths. That includes the AWS GovCloud (US) partition where it is required.
Safe defaults, visible settings
Agent capabilities start switched off. Turning one on is a deliberate act with a confirmation, so a workspace is never more capable than someone intended it to be.
- ✓ File editsapply_patch, write_file, notebook_edit, and saved plans.
- ✓ Shell commandsrun_command, PowerShell, background jobs, and /sandbox.
- ✓ MCP serversStart stdio/HTTP/WebSocket servers and their tools and prompts.
- ✓ Browser automationDrive a local Chrome/Edge over CDP.
Agent tools can read and modify files, run commands, and start local servers. Only enable for projects you trust.
$ /report Route Planner 2.4 readiness write document docs/route-planner-2.4-readiness.pdf
Files3
- route-planner-2.4-readiness.pdf 3.31 KB
- q3-proposal-summary.docx 18 KB
- depot-readiness.xlsx 24 KB
Saved into the working folder. PDF, Word, Excel and Markdown.
Beyond engineering
The same privacy, for the rest of the business
Contracts, proposals, pricing models and campaign plans are exactly the material an organization cannot paste into a public assistant. Topsoil writes real files (PDF reports, Word memos, Excel workbooks and Markdown) from a private library of your own documents.
- Proposal and contract work: draft, compare and summarize against prior awards and templates held in your own library.
- Marketing: briefs, campaign plans and positioning that can reference confidential roadmap material.
- Analysis: retrieve from PDFs, Office files and scanned documents, and get a spreadsheet or a report back.
- Shared projects and channels running in your own account, so a team can work on one thread without the files leaving it.
The terminal
A terminal AI agent for build servers, bastions and locked-down hosts
On a build server, a jump host or a locked-down workstation, a desktop application is not an option. The TUI is the full agent in a terminal, with the same models, permission gates and project context.
Every file it changed, as a diff. Each file is laid out line by line with syntax highlighting, so the review starts from the code before anyone opens a pull request. See the full walkthrough
At a glance
Topsoil features at a glance
| Capability | What it means in practice |
|---|---|
| Private inference | Requests go to the Bedrock, custom HTTPS or gateway endpoint you configure in your own AWS account. Your account and IAM policy decide where the work goes. |
| Model selection | The Amazon Bedrock models your account enables, grouped by vendor and selectable per session, project or workflow stage. The list differs by region. |
| Coding agent | Read and edit files, run approved commands, use Git, language servers, browser, web and document tools. |
| Plan-first governance | Read-only plan mode, tool risk classes, path protections, command allowlists, explicit approvals. |
| Project context | Projects, attachments and a per-identity context library backed by your own storage. |
| Local semantic search | An optional local index of your workspace that honours .gitignore and is built without uploading source. |
| Documents | PDF reports, Word memos, Excel workbooks, Markdown and rendered diagrams, saved into your working folder. |
| Integrations | MCP servers over stdio, HTTP or WebSocket, enabled per project, with secrets kept host-side. |
| Delivery | /task, /goal and /loop: backlog to pull request through isolated worktrees, with a verification step that must prove the test fails first. GitHub, self-managed GitLab, or Jira and Bitbucket Data Center inside your network. Topsoil does not approve or merge. |
| Worktrees and review | topsoil -w works in its own git worktree so your checkout stays clean; ask for a review and findings come back anchored to file and line before the pull request. |
| Pinned model credentials | The AWS profile you configure is the one Bedrock calls use, whatever else is exported in the shell, including role and credential_process profiles. |
| Session hygiene | New sessions start fresh with a hint to resume; topsoil sessions scrub redacts secrets saved before redaction covered them, in terminal sessions and the desktop databases. |
| Slack status | Optional, off by default: one-line task and budget status to a Slack channel. Metadata only. It never sends code, diffs, prompts or paths. |
| Spend limits | Per-run budgets, and monthly user and account budgets set by your admin in your own bucket, checked before each model call. An AWS budget action you deploy is the backstop. |
| Scheduled runs | Exit codes for finished, failed, did not start and budget stop, so a scheduler or CI job can act on the result. |
| Collaboration | Shared projects and channels running in your account, sharing the stream and selected files. Credentials, terminals and raw tool output are never shared. |
| Surfaces | A cross-platform terminal client, a VS Code context bridge, native desktop for macOS and Windows, and a customer-hosted web option. |
| Deployment | Your cloud account and region, including isolated and government partitions. Nothing is hosted by us. |
Topsoil is powered by Amazon Bedrock and runs on AWS, including AWS GovCloud (US), in your account and under your controls.
Questions
Frequently asked questions
Where does Topsoil run?
In your own AWS account, on Amazon Bedrock, in commercial regions or AWS GovCloud (US). Nothing is hosted by us and there is no Topsoil-operated service in the request path: inference runs through the Bedrock, custom HTTPS or gateway endpoint you configure, inside the account, region and network you already control.
Does our source code leave our environment?
No. Prompts, source, documents and output stay inside the account, region and network path you configure. There is no prompt store outside your environment, and no analytics pipeline, usage beacon or telemetry stream. That comes from the architecture, so there is no setting to turn it back on.
What does data sovereignty mean with Topsoil?
Your code, prompts and outputs stay in infrastructure you control: the machine that runs Topsoil, and the AWS account and Region you choose on Amazon Bedrock, commercial or AWS GovCloud (US). Topsoil runs no hosted service in the path. The model you pick matters too: a cross-Region inference profile can process a request in another Region within its geography, so choose a model ID that fits your residency rules. See the data sovereignty guide.
Which AI models can we use?
The Amazon Bedrock models your account enables. The list differs by region, including AWS GovCloud (US). Providers include OpenAI, Anthropic, xAI and Meta, among others your account turns on. AWS publishes which models are available in each region.
How do engineers stay in control of changes?
Nothing happens without your decision. Plan mode stays read-only until you approve, every edit arrives as a diff, and commands run only under allowlists and path protections. Agent capabilities start switched off. A handed-over task runs on its own branch in an isolated git worktree and comes back as a pull request or merge request for a person to review; Topsoil does not approve or merge it.
Where do engineers use it?
In a terminal client for macOS, Windows and Linux; a VS Code extension that lends the terminal agent your editor context and opens a diff of what it wrote; and a desktop app for macOS and Windows. It is the same product, with the same models and the same governance, in all of them.
Can it run over a private network path?
Yes. Topsoil works over public endpoints, custom HTTPS endpoints, or reachable PrivateLink and VPC paths, including the AWS GovCloud (US) partition. It needs a reachable Bedrock, custom HTTPS or gateway endpoint. Authentication uses IAM roles, AWS profiles, IAM Identity Center (SSO) or a gateway bearer token, and access, logging and retention stay under your existing policy.
Is Topsoil an alternative to GitHub Copilot?
For teams that want a private AI platform for engineering inside their own AWS account, yes. Topsoil is a self-hosted alternative to SaaS coding assistants: it runs on Amazon Bedrock in your account for software, infrastructure as code and review, with no vendor service in the request path. See the fair comparison.
Do we need to run GPUs or a model server?
No. Topsoil uses the models Amazon Bedrock serves in your account, so there are no model servers or GPUs to operate, and a new model is available when you enable it.
Does Amazon Bedrock use our prompts to train models?
AWS states that Bedrock inputs and outputs are never shared with model providers or used to train base models (AWS). Topsoil keeps no copy of its own: there is no Topsoil prompt store.
How do we request a pilot?
Use the form below, or email hello@thazen.com. A pilot runs against your own codebase, with your models and your material, and ends with findings you can act on. Requests go straight to the Topsoil team, who usually follow up within one business day.
Get started
Try it on 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.