VLR AI Harness for VS Code

Get multi-agent help with your code, or run autonomous repository improvements—all from VS Code.

Copilot Chat — choose VLR AI Harness in Copilot Chat to ask questions and work on files in your open workspace.
VLR Dream Studio — plan and carry out repository improvements with isolated workers, validation, and local merges.

The two modes have independent models, settings, and stop controls. A VLR subscription is optional; other available model providers work independently.

Get started

Prebuilt Marketplace packages include the local runtime for Windows x64/ARM64, macOS Intel/Apple silicon, and Linux x64/ARM64. Choose the matching platform if installing a VSIX directly.

Use VS Code 1.102 or newer. Install VLR AI Harness from the Extensions view if available, or choose Extensions: Install from VSIX… from the Command Palette and select the package for your operating system and architecture.
Open your project and save any pending editor changes. The installed platform package includes the local runtime; you do not need the .NET SDK just to launch it. Your project may still need its own build and test tools.
Open VLR AI Harness in the Activity Bar. Under Copilot Chat, choose the orchestrator (main agent) and default subagent models, then select Save Chat Settings.
Open Copilot Chat, select VLR AI Harness in its model picker, and describe your task.

Add a retry policy to the API client and cover it with tests.

Chat follows the active file's workspace folder, or the first open folder when no editor selects one. It does not require Git or create isolated worktrees. Cancel the Chat request to stop it.

Support development

Use the Sponsor button on the Marketplace listing to support VLR Creations.

Choose your models

Open a model card to browse or search the available providers. Use input-capacity filters, Favorites, or Recently used to narrow the list. Starring a model does not assign it to an agent.

Under Per-agent models, optionally choose different models for Planner, Explorer, Librarian, Designer, Worker, and Reviewer. Use default Chat subagent model clears an override. Select Save Chat Settings after changing any Chat assignment.

Dream Studio has its own supervisor and worker selections. Use Save Studio Models to save them; neither mode inherits the other's choices. Unavailable saved models remain visible so you can choose replacements.

Model access depends on your installed providers and may require sign-in, consent, subscriptions, connectivity, or available quota. Options include GitHub Copilot, Ollama, OpenAI, Anthropic, and the optional VLR AI provider, depending on your configuration. Choose a tool-capable Chat orchestrator with enough input capacity for your task.

Chat displays agent activity, public responses, and a final answer. Response usage totals can include multiple agent calls and may be marked estimated or partial. These totals are not the current context-window size or a billing statement; provider billing and usage policies still apply.

VLR Dream Studio

Dream Studio requires a trusted Git workspace and the tools needed to build and test your project.

Expand VLR Dream Studio and use Choose Repository if needed.
Choose the supervisor and worker models, then select Save Studio Models.
Review Run Settings, including workers, cycles, and change intensity.
Choose Run Dream Studio for autonomous improvements. To follow an existing Markdown or text plan, expand Plan (optional), choose Import Plan, then Run / Resume Plan.

Run Dream Studio ignores the selected plan. Plan runs use saved checkpoints and are not limited by the normal cycle cap.

Workers make changes in isolated Git worktrees. Candidates go through validation before local merges into the target branch; Dream Studio never pushes to a remote. Review changes and validation results before publishing them. AI-generated changes and passing checks do not guarantee correctness.

Use the live activity cards to follow a run, expand Output & diagnostics for the output console, choose Stop to cancel, and View History to review recorded work. Collapsing a section does not stop a run. Studio's controls do not stop Copilot Chat.

Advanced options

Dream Studio controls

Run Settings provides worker and cycle counts and change intensity. Expand its nested Advanced section for:

Target branch and worker branch templates.
Prompt, rules, and verification files; use Open to edit them and save before running.
Additional executable validation checks, supplied as a JSON array, for example: [{"Executable":"dotnet","Arguments":"test --no-build"}].
Before-run and after-success commands or scripts.

Plan (optional) contains plan import/resume controls; Output & diagnostics contains the console and automatic output scrolling. Both sections start collapsed.

Checks and hooks execute commands on your machine. Use only files and commands you trust.

MCP and settings

Compatible MCP clients, including Copilot agent workflows, can use the vlr_ai_improve tool for Dream Studio work. MCP has separate direct-provider settings; saving sidebar models does not configure those credentials or models.

Search for VLR AI Harness in VS Code Settings for model, repository, runtime, validation, hook, and licensing options. Runtime overrides are optional; most users should keep the bundled runtime. If your installation requests a license, use VLR AI Harness: Set License Key.

Optional VLR AI account

Expand VLR AI Subscription and choose Connect VLR AI to sign in through VS Code's email and password prompts. You can instead provide an existing API key under Connection options.

The extension is free to install from the Marketplace. Marketplace listings support Free and Trial labels, but not a numeric extension price or Marketplace checkout. VLR AI subscriptions are sold separately on the VLR website at the prices shown there.

Signing in creates a dedicated, revocable extension API key. The key and connection metadata are kept in VS Code SecretStorage; your password, access token, and refresh token are not saved or sent to the sidebar. Do not paste credentials into chat or workspace settings.

Disconnect removes the saved connection from this device only. To disable the key everywhere it is used, choose Manage / Revoke API Keys and revoke it on the website. Manage existing subscriptions through your account on the website.

After connecting, select Auto, Fast, or Deep under VLR AI in the model library and save each desired mode or agent assignment. Connecting does not change existing model choices.

Access, billing, and limits

Basic alone does not include external API access for this provider. An active Max subscription or explicitly enabled API access is required.
Purchase buttons open the website for review and confirmation; clicking a button or installing the extension does not purchase a subscription.
Signup creates a recurring monthly, invoice-based subscription, not an immediate card checkout. The selected chat seat is billed together with included billable image-generation and prediction seats. Invoices are due within 30 days. Review current prices, the full monthly subtotal, and applicable taxes or fees before confirming.
VLR AI input budgets are API character-budget estimates, not the underlying model's context-window capacity. Request/output limits, rate limits, monthly budgets, account status, and service availability apply. Large Copilot Agent contexts may need a higher-capacity provider.
Available models and streaming/tool support depend on the service version. If purchase or model access is unavailable, follow the displayed error or use another available provider.

Workspace access and privacy

The harness runs locally, but local execution does not mean all model processing stays on your device. Requests and relevant workspace content are sent to your selected model provider. Review that provider's privacy, retention, and usage policies before sharing sensitive code.

Chat's file tools work within the selected workspace folder; attachments do not grant access outside it. Tools read and change files on disk, so save editor changes before asking for edits. Dream Studio can run project checks and configured scripts. Trust only repositories and commands you understand, and review AI-generated edits before committing.

Help and diagnostics

Run VLR AI Harness: Show Diagnostics from the Command Palette for runtime logs. If models are missing or unavailable, check provider sign-in, consent, quota, and model selections, then refresh and save an available model. For VLR AI access errors, check your subscription/API entitlement and key status.

When requesting support, include your extension version, VS Code version, operating system, affected mode, and the relevant error. Review diagnostics before sharing them and remove secrets or private project content.

License

This extension is provided under a commercial license. See the included LICENSE.txt for terms.