Microsoft Amplifier

What Is Microsoft Amplifier?

Amplifier is Microsoft's open-source research project exploring modular, extensible AI-assisted development. Currently a CLI tool, it's designed as a platform where the CLI is just one interface — with web, mobile, voice, and collaborative AI-to-AI interfaces on the roadmap.^[1]

The project is explicit about what it is: a research demonstrator, not a product. Microsoft is performing "active exploration in the open" and hasn't built safety systems yet.^[1] This honesty is refreshing — and important context for anyone evaluating it.

Previously built on top of Claude Code, Amplifier has since been rewritten with its own modular kernel. The old Claude Code version lives on the amplifier-claude branch.^[1]

Status as of June 11, 2026: active development (last push June 9, 2026), ~3.1K stars, 255 forks, MIT license, not archived — and still zero tagged releases since the repo's creation in September 2025.^[2] Microsoft now maintains an official docs site at microsoft.github.io/amplifier-docs.^[3]

Architecture: Bundles, Modules, and Agents

Amplifier's core idea is composability through bundles. A bundle is a configuration package that defines:^[3]

• Providers — which AI model to use (Anthropic, OpenAI, Azure, Ollama)
• Tools — filesystem, bash, web, search, task delegation
• Agents — specialized AI personas for focused tasks
• Behaviors — logging, redaction, streaming UI, todo tracking

The default foundation bundle ships with 14 specialized agents:^[1]

Bundles can be installed from git repos, switched on the fly, or scoped per-project vs. per-user. The amplifier bundle add command pulls external bundles, and amplifier bundle use switches context.

The Kernel

The architecture splits into three layers:^[4]

1. amplifier-core — Ultra-thin kernel (~2,600 lines) providing module protocols, session lifecycle, and hooks
2. amplifier-foundation — Bundle composition library + the default foundation bundle
3. amplifier-app-cli — The reference CLI implementation

The docs describe a Linux-inspired kernel philosophy: "The center stays still so the edges can move fast."^[3] This separation means the platform can support entirely different frontends while sharing the same module system.

Multi-Provider Support

Amplifier supports four AI providers:^[1]

• Anthropic Claude — Primary, most tested (Sonnet 4.5, Opus 4.6, Haiku 4.5)
• OpenAI — GPT-5.2, GPT-5.2-Pro, GPT-5.1-Codex
• Azure OpenAI — Enterprise path with managed identity and az login support
• Ollama — Local inference with llama3, codellama, mistral

Switching is a single command: amplifier provider use openai --model gpt-5.2. Provider config can be scoped to local (just you) or project (team-wide).

The Azure OpenAI support is notable — it uses DefaultAzureCredential, which means it works with az login locally and managed identity in production. This is the enterprise-friendly path that most open-source agents skip.

Session Persistence

Sessions are automatically saved and project-scoped. Working in /home/user/myapp shows only that project's sessions. amplifier continue picks up the last session; amplifier session resume <id> resumes a specific one.^[1]

This project-scoping is a nice design choice — it mirrors how developers actually context-switch between codebases.

Self-Improvement Patterns

One of Amplifier's more interesting claims: "Use Amplifier to create new Amplifier modules." The system is designed to be self-extending — agents can build new modules, agents, and bundles. This is the "AI builds AI" part of their vision.^[1]

The modular architecture makes this more plausible than it sounds. Since modules follow defined protocols and bundles are just configuration packages, generating new ones is a structured problem rather than open-ended code generation.

One caveat worth flagging: the metacognitive-recipe and DISCOVERIES.md learning patterns that made the original Claude Code-era Amplifier notable no longer appear in the current README or docs — they live on in the legacy amplifier-claude branch, while the rewrite's self-improvement story is currently limited to self-extension through modules.^[1][3]

Why It Matters

Microsoft releasing a research-stage agent framework in the open is significant for a few reasons:

1. Honest positioning — Calling it a research demonstrator sets correct expectations. Most tools in this space overpromise.

2. Bundle composability — The idea that agent configurations should be packaged, shared, and composed like libraries is compelling. Most agent frameworks treat configuration as monolithic.

3. Multi-provider from day one — Supporting Anthropic, OpenAI, Azure, and local models means teams aren't locked in. The Azure path is particularly important for enterprise adoption.

4. Platform, not just a CLI — The kernel/foundation/app separation suggests Microsoft is thinking about this as infrastructure, not just a developer tool.

Cautions

• Not accepting external contributions — still the policy as of June 2026, though the team says it is "actively working toward opening this up"^[1]
• No releases, ever — zero tagged releases or published packages; install is uv tool install straight from the git repo^[2]
• Safety systems not built — they're explicit about this^[1]
• Anthropic-first — other providers have "rough edges"
• Documentation incomplete — the new docs site helps, but coverage is still catching up^[3]
• Flat adoption — ~3.1K stars in June 2026, barely moved from ~3K in February; development is active but the community isn't growing with it^[2]

What Developers Say

The only substantive public discussion remains the October 2025 Hacker News thread (262 points) — which covered the original Claude Code-based version, before the modular-kernel rewrite. No notable HN or Reddit discussion of the rewritten Amplifier was found as of June 11, 2026.^[5]

The thread skewed skeptical:

"Again this 'supercharging' nonsense? Maybe in Satiyas confabulated AI-powered universe, but not in the real world I am afraid..." — hansmayer, Hacker News^[5]

"This project was in part written by Claude, so for better or worse I think we're at least 3 levels deep here (AI-written code which directs an AI to direct other AIs to write code)." — jsheard, Hacker News^[5]

"'Eat your own dog slop' isn't bad practice, though. Some people in the organization will experience the limitations and some will learn." — rectang, Hacker News^[5]

"There are many many people who want better AI coding tools, myself included. It might or might not fail, but there is a clear and strong opportunity here, that it would be foolish of any large tech company to not pursue." — falcor84, Hacker News^[5]

The silence since the rewrite is itself a signal: Microsoft is iterating in the open, but almost nobody outside Microsoft is publicly using or discussing it.

Pricing & Licensing

Licensing model: MIT — free for commercial and personal use.^[2]

Hidden costs: You bring your own model access — Anthropic/OpenAI API keys, an Azure OpenAI deployment, or local compute for Ollama.

The Tembo Angle

Amplifier's bundle system is philosophically aligned with agent orchestration — composable, modular, multi-provider. The key difference: Amplifier bundles are static configuration packages, while true orchestration involves dynamic routing, cost optimization, and runtime decisions about which agent handles what. Amplifier is a single-agent system with delegation; orchestration platforms manage fleets.

Worth watching as Microsoft iterates. The modular kernel could become infrastructure that other tools build on — or it could remain a research artifact. The "not currently accepting contributions" stance suggests Microsoft isn't sure yet either.

Bottom Line

Amplifier is the rare big-company agent project that under-promises: a research demonstrator with a genuinely interesting kernel/bundle architecture, honest disclaimers, and steady commits — but nine months in, it has no releases, no open contribution path, and a community that stopped talking about it after the October 2025 launch thread.^[2][5]

Recommended for: Researchers and framework authors studying modular agent architecture, and Azure-centric teams prototyping with managed-identity model access.

Not recommended for: Production use of any kind (Microsoft says so itself), or anyone who needs the metacognitive-recipe self-improvement patterns — those now live only on the legacy branch.

Outlook: Development is active and the kernel design is sound, but flat stars and zero post-rewrite discussion suggest Amplifier is still an internal research vehicle that happens to be public. The signal to watch is whether Microsoft opens contributions and tags a release — until then, it's an architecture worth reading, not a tool worth adopting.

---

Research by Ry Walker Research • methodology