Personal assistance for the Windows workday.

The work between the work.

Starting a task means finding a project, opening its editor and references, recovering a decision and setting up the next action. These repeated transitions interrupt the work people sat down to do.

Diamond is being built for people who already work across desktop tools. The first beta is deliberately focused on three recurring jobs:

1. Prepare a workspace. Resolve a named project and open its editor, folder and references.
2. Repeat a personal routine. Save a familiar setup under a phrase, then reuse it with another project.
3. Move a meeting forward. Recover a decision and prepare a reviewable follow-up with a source passage.

Cloud understanding. Local execution.

Voice and text enter one orchestrator. The host backend handles authentication, provider credentials, model requests and usage. A bounded task graph returns to Diamond Core on the user’s Windows PC.

Core resolves names against real local resources, checks Trust Center access and runs registered tools. Task state keeps completed results separate from unfinished work; uncertain outcomes are not blindly replayed.

The implemented speech layer uses Deepgram; the current Computer Use baseline uses GPT-6 Luna. Claude is the next proposed provider evaluation, not an integration already deployed in this build.

Claude Model Routing

A two-tier Computer Use routing design separates routine interactions from complex, context-dependent operations. The two roles align with the Claude model families and share the same tool contracts and local permission checks.

The routing policy assigns simple, bounded actions to the Haiku role and complex operations to the Sonnet role. Task complexity, step dependencies, interface ambiguity and recovery needs determine the route; escalation preserves completed work and the remaining task context.

A shared execution boundary keeps resource resolution, tool validation, permission checks and result verification in Diamond Core. Model reasoning does not replace local execution controls. Routing quality is assessed through task completion, user corrections, latency and cost per verified result.

A product with a testable foundation.

The desktop client, host backend, local resource index, routines, task lifecycle, memory and meeting workflows are implemented. The site shows the actual English interface using an isolated demonstration profile.

The full offline baseline on 7 October 2026 recorded 2,946 passed checks, 2 skipped and no failures. It did not call paid providers or perform physical Windows actions. Live task completion and separate-PC release validation are the next gates, not claims derived from that test count.

The current release candidate also needs signing and distribution validation before the invited cohort receives it.

Support that turns an evaluation into a product.

We are seeking support through Claude for Startups for a focused integration and beta cycle:

• API credits for measured product work. An initial $1,000 would support comparative workflow runs, regression cases and beta cost measurement, with per-run budgets and usage tracking.
• Claude Code for development. Use it for adapter implementation, tool-contract tests, error analysis and release hardening.
• Applied AI office hours. Seek feedback on structured tool requests, source-grounded synthesis, prompt caching and the latency–cost trade-off.

These are requested uses of support, not an award or partnership announcement. Available benefits follow the current program and approval terms.

The proposed first six months.

1. Establish the Claude baseline. Verify the adapter, run held-out comparisons and improve the failures before expanding scope.
2. Make the first two PCs reliable. Validate host/client isolation, voice, cancellation, reconnect and sleep/wake through real recurring workflows.
3. Open a focused invited beta. Target 10–20 testers after the release gates pass. Measure completion, time to a useful result, user corrections, return use and cost per completed task.

The planned business model is a subscription with included usage. Beta retention and measured service costs will guide limits and pricing. The cohort is a target, not an existing customer count.