- “Local” products secretly depend on a central GPU
- Inference and training are advertised as the same thing
- Hardware fit is guessed from model name alone
private ai workflow hub
Private and local-first AI workflows
Run compatible text models from user-owned hardware, keep local connection details in the browser, and train only through explicit GoldCap actions.
- Keep endpoint and pairing token device-local
- Never expose the companion to a public interface
- Separate data preparation, process completion, adapter activation, and quality
- Recheck hardware before changing model size or context
One flow, bounded owners.
| Stage | Purpose | Suggested ownership |
|---|---|---|
| Hardware fit | Estimate memory, storage, runtime, and training boundaries | User + hardware checker |
| Pairing | Connect the browser to the loopback companion | Local Companion |
| Inference | Use the local model in any entitled box | Any assigned primary role |
| Data preparation | Build an explicitly approved local dataset | GoldCap |
| Training and test | Launch a compatible toolkit and validate the result separately | GoldCap + user |
Authored recipes
Start from a tested structure.
solo · Builder 5
Inspect + download →Private local-first document review
Assign bounded review responsibilities to local-model boxes without routing the source to a cloud provider.
Roles + task guides
Understand each responsibility.
primary role
Open role →Reviewer
Evaluates a completed deliverable against explicit requirements and issues a clear disposition.
primary role
Open role →Summarizer
Compresses supplied material while preserving the decisions, evidence, and boundaries that matter.
support role
Open role →Context keeper
Protects the shared objective, canon, decisions, and accepted handoffs across a long workflow.
support role
Open role →Fact-checker
Challenges factual claims and source alignment inside another role’s work.
task guide
Follow guide →Install the Local Companion on Windows
Start the loopback companion and pair it with the same computer’s browser.
task guide
Follow guide →Install the Local Companion on macOS or Linux
Start and pair the device-local bridge without creating a site-owned model endpoint.
task guide
Follow guide →Use a local model in any entitled box
Route normal box inference through the paired user-owned model; reserve GoldCap for training work.
task guide
Follow guide →Prepare GoldCap training data
Create a canonical local dataset from explicitly approved pairs or saved-workflow scope.