An interview-practice application built with Microsoft Agent Framework, models hosted in Microsoft Foundry, MCP tools, and Aspire. It collects a resume and job description, asks behavioural and technical questions, gives feedback, and saves an interview summary.
Learn by building: The workshop guides you through constructing the application from a cloud-free starter. Fifteen lessons start with tool setup and a completed-example run in Chapter 0. Then keep working in a separate starter: create an agent, stream its replies, add tools and records, read documents, and build specialist handoffs. The starter supplies the UI, repository, and a source-derived WorkshopHosting.cs helper for model authentication, client setup, optional Copilot support, and DevUI. It starts cloud-free. In the first-agent lesson you write the ChatClientAgent constructor and coaching instructions, then activate the supplied hosting. Core orchestration edits stay in root apphost.cs; the capstone's explicit support patch prepares the completed deployment entry point.
Run the finished application: This repository contains the completed application. Clone it, configure it, and run it to see the interview coach in action. Use the reference documentation below to understand the architecture, configuration, and data contracts. You don't need the workshop to run the finished version.
For an exact setting or contract, use the architecture, configuration, agent modes, and session data references.
The browser connects to a server-interactive Blazor UI. The WebUI server sends chat messages to the .NET agent service through AG-UI. Agent Framework runs either one coach or a handoff workflow with triage, receptionist, behavioural interviewer, technical interviewer, and summariser roles. The workflow transfers control between roles that share the configured model deployment.
InterviewData MCP exposes the Cosmos-backed interview repository. MarkItDown MCP extracts document text. The model requests tool calls, and application code executes them.
The agents run in the application's .NET process. Foundry hosts the model. Aspire starts the local services and supplies their connections. The UI message list and session ID live in the WebUI's server-side Blazor circuit. Refreshing creates a new circuit and session; earlier interview records remain available through InterviewData.
The commands below run the completed application in this checkout. Both root apphost.cs and src/InterviewCoach.AppHost contain the completed resource graph here. Workshop downloads have their own staged scaffold; their project-based AppHost stays in its starter state until capstone finalization. See entry points.
For the pinned, guided version, start with Chapter 0. To run the current checkout, install .NET 10, the Aspire CLI, a compatible container engine, and the Azure CLI.
A local run can provision billable Azure resources. The default AppHost declares a Foundry resource and model deployment. Before starting, confirm the subscription, region, model availability, quota, provisioning permissions, and cleanup owner. See Foundry configuration. Use fictional interview data and a development environment you own.
git clone https://github.com/codemillmatt/interview-coach-agent-framework.git
cd interview-coach-agent-framework
az login
az account show --query "{subscription:name, subscriptionId:id, tenant:tenantId}" --output tableReview apphost.settings.json before the next command. It selects MicrosoftFoundry and HandOff by default and requests a specific model version and capacity. The agent uses DefaultAzureCredential; do not put API keys in source.
From the repository root, with the container engine running:
aspire start --apphost ./apphost.csOpen the dashboard URL printed by Aspire, inspect the resources, and use the webui endpoint when its dependencies are ready. Try the fictional samples and inspect the saved record as described in application usage.
The application also implements GitHub Copilot as an optional provider and single-agent mode as a simpler baseline.
From the folder that started the AppHost:
aspire stop --apphost ./apphost.csCloud resources remain until cleanup. Keep an inventory of the resources created by each run. If you later use the optional Container Apps deployment, its azd environment has a separate cleanup scope. Follow the cleanup reference and remove only resources you own.
This is a learning sample. Review authentication, per-user record access, exposed development endpoints, uploads, and retention before using real interview data or opening it to other users.
The static Astro/Starlight website is separate from the .NET application. With Node 24 installed:
cd workshop
npm ci
npm run devThe build generates source-backed checkpoints and imports the canonical reference from docs/ without Azure credentials. See workshop authoring for the content and checkpoint contracts.
Contributing explains how to report issues and submit changes. The project uses the MIT license.