microsoft/ai-agents-for-beginners

Source
GitHub
First trending
Category
Agents
GitHub stars
80,059
Main language
Jupyter Notebook
Website
aka.ms/ai-agents-beginners (opens in a new tab)

This page introduces an external open-source repository. It is not an HDATF product.

microsoft/ai-agents-for-beginners

What it does

This is a beginner course for developers learning to build AI agents, rather than a finished business application. Its lessons combine explanations, videos and executable examples covering tool use, retrieval, planning, multiple agents, memory and security. Learners supply questions, tool definitions and sample data, then inspect agent responses and structured results. The Python tool-use notebook shows how typed functions become callable tools and how a schema constrains the returned object. It also distinguishes automatically executed tools from tools that require human approval. The main examples use Microsoft Agent Framework and Microsoft Foundry, with Python 3.12 or later, an Azure subscription, authentication and a deployed model. Some examples support alternative providers or local models, but their configuration and API capabilities differ. The travel example uses fixed destination and flight data, so it demonstrates a pattern rather than a working booking integration or production safety guarantee.

License

MIT Permissive. Commercial use and changes are allowed if the copyright notice is kept.

More in this category

  • open-jarvis/OpenJarvis
    OpenJarvis is a local-first personal AI framework for users and researchers who want more control over inference and agent operation. Users provide questions, indexed documents, agent settings and optionally connected account data. Its presets cover simple chat, cited research, code assistance and scheduled monitoring, rather than one fixed assistant workflow. The research guide describes splitting files into chunks, storing them in SQLite with full-text search and repeatedly retrieving evidence for an answer. Local inference can use Ollama, while the cloud engine source also implements external provider routing, so local-first does not guarantee an offline or local-only deployment. Web research and account connectors can introduce additional external services. The README describes evaluation of quality alongside latency, energy and cost, plus learning from local traces, but those claims were not reproduced here. The project requires Python 3.11 or newer and an appropriate model runtime, and answer quality, resource use and privacy boundaries depend on the selected model, tools and configuration.
  • yetanotherbrowser/yab
    Yab is a browser for macOS with a companion iPhone app and agents that work in their own tabs. WebKit is the default page engine, while Apple silicon Macs can select Chromium per tab or site. A shared page interface hides engine differences, and Chromium pages run in a separate CEF helper process. Agents can use existing site sign-ins, but private tabs are excluded from their visibility and actions. Actions such as sending messages, uploading files and paying can require a person's approval, with eligible requests shared to other synchronized devices. Approval records bind answers to a request digest and expiry, rather than treating any later yes as valid. Optional synchronization stores encrypted records on the server, but pages read by an agent are sent to its model provider under that provider's terms. The README requires macOS 15.4 or later, and the architecture notes that some WebKit-only features have no Chromium path.
  • ysunyang4/Nexus
    Nexus connects external AI clients to a local execution system through MCP, a protocol for calling tools. The repository retains LocalBridge names for internal packages and services. A desktop control center, HTTP server, and authenticated runner divide configuration, requests, and local execution. Its tools cover project files, Git, commands, browser operations, memory, and knowledge. Access settings distinguish read-only from read-write projects and separately control execution. The implementation separates MCP client, runner, and local management token types, while authorization must still be checked for each ingress route. An action ledger records execution and verification states for recovery, without establishing that every external side effect is reversible. The inspected decision-provider type offers disabled or Laya advisory operation, not arbitrary selection among hosted chat models. Documentation differs across the core server, bridge, and cloud gateway, and full-control mode expands the risk of local changes.
  • homeassistant-ai/ha-mcp
    HA-MCP exposes Home Assistant management operations to AI clients through MCP, the protocol used to describe and invoke application tools. An AI client turns a user's request into structured tool arguments, and the server returns device states, histories, configuration data or operation results from the connected Home Assistant instance. Its tool registry covers device control as well as automation, script and dashboard editing and diagnosis from logs and execution traces. The inspected server initializes a FastMCP tool surface with visibility filters, optional search-based discovery and per-call enforcement rather than relying only on a model instruction to be careful. The recommended in-process integration requires a compatible Home Assistant installation, with the documented server setup floor at Home Assistant 2026.8.0, and an administrator account's access credential. Local and remote connection options differ, and a secret connection URL can itself be a credential that must be protected. Read-only connections hide and block writes, but the same credential can still access the normal endpoint, so this mode is not a separate credential-level permission boundary. Direct file and YAML editing are opt-in beta features, and backups and configuration checks do not eliminate semantic errors or the need for independent recovery access. Running the server locally does not guarantee local AI processing, because the selected MCP client can process Home Assistant data under its own privacy policy. This is a Home Assistant administration tool, not evidence of compatibility with industrial control systems or a guarantee that generated automation is safe.
  • Ebony-Vinyl/EAC-Desktop
    EAC-Desktop is an application that packages DeepSeek Harness as a desktop workspace. It targets users who want conversations, project-file work, tool execution and plugin management around a project directory. Task requests, project files and model configuration feed a workflow in which Harness runs conversations and tools and the interface exposes file changes and results. Its documentation separates a Tauri window and tray shell, a Node sidecar for startup and environment orchestration, and a Harness agent kernel. Described capabilities include reviewing and reverting file changes, conversation compaction, model configuration and plugin-management entry points. Distribution packages prepare runtime resources, but model calls depend on the configured provider, while plugin activation and extra dependencies vary by package. The inspected environment adapter calls DPX environment creation, path and runtime-variable APIs to separate default data locations by release channel. This is not a filesystem sandbox: absolute-path access still depends on operating-system permissions, and third-party component licenses require separate checking.

Only repositories in the ranked Trendshift lists are included, and the lists are used only to find candidates. We do not copy their ranks. Descriptions, licenses and star counts come from each GitHub repository. The notes are our own reading. We have not tested these projects, and a place on a trending list does not prove quality.

View on GitHub (opens in a new tab)