repository·main·Indexed 25 days ago
https://github.com/microsoft/amplifierAn AI-powered modular development assistant for the command line. Amplifier uses a modular architecture where providers, tools, and behaviors are composed using 'bundles'. It supports multiple AI providers including Anthropic Claude, OpenAI, Azure OpenAI, and Ollama, and features session persistence, agent delegation, and an extensible module system for building specialized CLI tools and agents.
The assets/branding/ directory contains official Amplifier branding assets, including icons, favicons, and PWA assets. The directory structure is as follows:
icons/: App icons in various sizes and formats (PNG, ICNS, ICO).favicons/: Web favicons for browsers and iOS.pwa/: Progressive Web App icons.Amplifier uses a modular architecture consisting of core infrastructure, applications, documentation, libraries, and bundles.
Applications are user-facing tools that compose libraries and modules. Key applications include:
amplifier-app-cli via uv tool install.amplifier-core and amplifier-foundation over REST and SSE, allowing you to drive sessions from any language or framework.amplifier-agent.amplifier-opencode) that configures the opencode TUI using amplifier-agent's OpenAI-compatible HTTP face.amplifier_local adapter on the Amplifier engine.amplifierd.amplifierd using the OpenAI Realtime API.Libraries are used by applications (like amplifier-app-cli) but are not used directly by runtime modules. Runtime modules only depend on amplifier-core.
Amplifier uses a Linux kernel model where amplifier-core acts as an ultra-thin kernel and everything else is implemented as modules.
Module Types:
The amplifier bundle serves as the primary entry point for the entire Amplifier ecosystem. It provides ecosystem overviews, getting started guidance, and access to the amplifier-expert agent for authoritative consultation.
Key components of the ecosystem include:
Bundles are capability packages that define what tools and agents are available to you. They are independent of the Provider (the AI service).
| Bundle | Purpose | Tools | Agents |
|---|---|---|---|
| foundation | Bare minimum | filesystem, bash | None |
| dev | Full development | base + web, search, task | zen-architect, bug-hunter, modular-builder, explorer, researcher |
| recipes | Multi-step workflows | base + task | Recipe execution agents |
| full | Demo/Showcase | Almost everything | Broad showcase |
Note: You can use the dev bundle with any provider (Anthropic, OpenAI, etc.).
Amplifier operates as a mechanism rather than a policy engine. The Kernel provides small, stable interfaces that modules implement.
Key core components include:
Policies (such as provider choice, logging detail, and privacy redaction) are handled by edge modules, while the core remains text-first and inspectable using JSON schemas and a canonical JSONL stream for observability.
The Amplifier ecosystem is organized into a modular hierarchy. Understanding where specific information or components reside helps you navigate the project and its dependencies:
amplifier): The main starting point for users. Contains getting started guides, ecosystem overviews, and links to all components. Use this for initial setup.amplifier-core): The ultra-thin core providing mechanisms like sessions, coordinators, and events. It defines core contracts like the Mount Plan and event taxonomies.amplifier-foundation): Provide APIs and primitives. amplifier-foundation is the primary library for bundle primitives (composition, validation, resolution), shared utilities (deep_merge, file I/O), and reference bundles.amplifier-module-provider-openai): Kernel extensions categorized into Context Managers, Hooks, Orchestrators, Providers, and Tools. They implement specific functionalities and are mounted via kernel contracts.amplifier-bundle-recipes): Composable configuration packages that combine providers, behaviors, agents, and context. They are consumed by applications via amplifier bundle use.amplifier-app-cli): End-user tools that consume libraries and modules. The CLI provides command references and specific implementation details.Amplifier is a modular AI agent framework based on a kernel-and-module philosophy. The architecture is divided into four main layers:
amplifier-core): An ultra-thin layer providing session lifecycle management, module loading, event systems, and coordinator infrastructure. It provides mechanisms, not policies.amplifier-foundation): The primary library for building applications, providing bundle primitives, reference behaviors, and shared utilities.anthropic, openai, azure, ollama).filesystem, bash, web, search).loop-basic, loop-streaming). Swapping this changes the entire execution loop.context-simple, context-persistent).logging, redaction).Modules are loaded dynamically at runtime based on your bundle configuration. You can define tools within a bundle using a YAML frontmatter block. This allows you to specify the module ID and its source (e.g., a git URL) to extend the agent's capabilities.
---
bundle:
name: my-bundle
version: 1.0.0
tools:
- module: tool-web
source: git+https://github.com/microsoft/amplifier-module-tool-web@main
- module: tool-custom
source: git+https://github.com/you/your-custom-tool@main
---
# My Bundle Instructions
Your system prompt here.All Amplifier modules must follow a consistent architectural pattern to ensure compatibility with the kernel. When building a module, follow these four steps:
mount(coordinator, config) function.coordinator.Tool, Provider, or Hook).For low-level kernel interfaces and protocols, refer to the amplifier-core repository.
In interactive chat mode, you can delegate tasks to specialized agents. These agents work in focused sub-sessions with specific capabilities.
Available Agents (in dev bundle):
zen-architect: System design and architecture reviews.bug-hunter: Systematic debugging and root cause analysis.modular-builder: Creating new, self-contained components/modules.researcher: Researching and synthesizing external information.explorer: Mapping local files and assets.How to use: Simply ask the agent to perform a task in chat:
> Delegate to zen-architect: Design a caching system> Use bug-hunter to find issues in src/main.py