01
Your code never leaves
Prompts, diffs, and source stay on a host you control. Federated learning does not upload the repo.
A Blue6 company
Keep your alpha.
Six principles for a private desktop coding agent. Open weights. Your code never leaves. In the app, it’s B6.
B6
WaitingOpenNew
Refactor auth middleware and open a PR. Do not send anything outside this folder.
Session-token check in
src/middleware/auth.ts, then a test run.- Edit
verifySession(req) - Run auth tests
✓✕
Describe what to build, fix, or explain… (/ for skills, ! for shell)
The six
Six principles. That’s the product.
Open-weight models. Private inference. Your edge stays yours. On the site it’s Alpha. In the dock, the model picker, and the adapter — B6.
01
Private inference
The agent cannot reach a public model API. Local or dedicated AWS Bedrock — if you host it, Alpha can use it.
02
Open weights
The model is a choice you own, not a vendor. Swap a qualified private endpoint. Alpha does not pick the brand of weights.
03
Your code stays
Prompts, diffs, and source stay on a host you control. There is no fallback onto someone else’s infra.
04
Consent is off
We do not train on your work by default. Federation is a switch in the product, not a clause in the footer.
05
Deltas, not files
If you opt in, traces are redacted on the device. Only adapter weight deltas federate. The repo never does.
06
Keep your alpha
Closed agents get smarter on your IP. Alpha inverts that. The more the team uses it, the better your private model gets at your stack.
Federated learning
Data never leaves. The model still gets smarter.
Closed agents get smarter on everyone else’s code — including yours. Alpha inverts that. Your source stays on a host you control. If you opt in, only adapter deltas federate. Usage compounds a private model without the IP leaving.
How it learns
- 01
Your code stays
Source never leaves the host.
- 02
Only deltas move
Tiny adapter updates, not files.
- 03
The model compounds
A private B6 adapter, trained on your work.
Consent is off by default. Raw source never federates. The adapter is B6.
02
Only deltas move
Consent is off by default. If you opt in, traces are redacted on-device. Secrets and paths stay local. Only adapter weight deltas federate.
03
The model compounds
Those deltas train a private B6 adapter in your tenancy. The more the team uses it, the better it gets at your stack — without the IP leaving.
Privacy first
Nothing leaves a host you control.
Other agents can optionally go local. Alpha cannot call a public model API. Prompts, code, and traces stay on your endpoint. Learning is opt-in, in-tenancy, and adapter-only.
Inference stays in the room.
LocalAWS Bedrock
Blocked
- ChatGPT / Claude APIs
- OpenRouter, Groq, Together
- Public Gemini / HF router
- Any hosted SaaS model you do not run
Allowed
- Local (Ollama, llama.cpp, MLX)
- Dedicated AWS Bedrock
No third-party model APIs
OpenAI, Anthropic, OpenRouter, Groq, hosted gateways — unreachable. There is no “just this once” fallback onto someone else’s infra.
Your weights, your box
Local runtimes or dedicated AWS Bedrock. If you host the weights, Alpha can call them. We do not.
Telemetry off
Sessions stay on the desktop unless you opt in. We do not train on your work by default. Consent is a product surface, not fine print.
You can watch the work
Thoughts, diffs, commands, and subagents render live. Trust comes from seeing what the agent did, not from a vendor claim.
First run
Open a folder. Keep the work here.
Built for people who are not going to maintain a terminal agent. End users never type an API key. Admins push endpoints once. Mac now, Windows on the same codebase.
Open a folder
The workspace is a directory on disk. The agent works where your repo already lives.
Point B6 at your host
Local or dedicated AWS Bedrock. If you host it, B6 can use it.
Let it run, then steer
Follow the live transcript. Send a correction mid-run. The session never leaves the host you control.
Private preview
Keep the code. Keep the alpha.
Private preview for teams that cannot send source to a public model — and will not accept an agent that never gets better at their repo.
Contact Alpha