Anthropic AI stack — what they build with and how

Anthropic's public-facing stack is mostly inferred from their hiring posts, GitHub org, and the rare engineering write-up [1][2]. The Claude.ai web app is a React frontend with a Tailwind-flavoured component system; the Console (developer surface) uses the same framework and shares the visual language [3].

Backend is a mix of Python and Rust. The Python API surface fronts the model serving infra; Rust shows up in roles for performance-critical inference paths and tooling [2]. They run their own inference fleet rather than wholly renting from cloud providers — the volume of GPU/TPU and "inference platform" job postings on their careers page through 2024-2025 is hard to read any other way [2].

The reference servers for MCP (Model Context Protocol) — the open standard they shipped in late 2024 — are TypeScript and live in the public anthropics GitHub org alongside SDKs in Python, Node, Go, Java, and Ruby [4]. Models are exposed via the Anthropic API directly, plus AWS Bedrock and Google Vertex AI as resale partners [5].

What's not public: the specifics of their RL training infrastructure, their internal evaluation harness, and the fine-grained data pipeline beyond the broad "curated and filtered web data" framing in their model cards [5]. None of this is hidden adversarially — it's just the kind of detail a frontier-model lab doesn't put in a blog post.

Anthropic is hiring aggressively across research, product, and platform. As of April 2026 the public careers page lists ~250 open roles split between San Francisco (HQ), London, Zurich, and remote [1]. The shape of the roster: roughly half research and research engineering, the other half product, GTM, and platform.

Notable concentrations: applied AI engineers (the "hands on a keyboard" agent-building roles), RL engineers, and a steady stream of trust & safety / policy hires [1][2]. They're also hiring inference engineers and "model performance" specialists, which lines up with the self-hosted inference signal in the stack section.

LinkedIn shows headcount roughly tripled across 2023-2025 and continues to grow at a measured pace; they aren't doing the public-spectacle hiring waves you see at OpenAI [3]. Compensation tracks frontier-lab norms — the leaked 2024 ranges put senior research scientists at $300k-$700k base plus equity [3].

If you're a solo founder studying who to hire — or who to compete with for talent — the salient pattern is that Anthropic's roles are unusually specific. "RL engineer focused on reward modelling" rather than "ML engineer". That specificity is itself a hiring signal: the team is large enough to specialise.

Anthropic ships the models. The Claude family — Haiku, Sonnet, Opus — covers the price/intelligence curve from $0.80/M input tokens (Haiku) to $15/M (Opus) as of April 2026 [1]. The 4.7 generation (current at writing) added the 1M-token context window on Opus and improved tool-use latency across the family.

What they ship around the models matters as much as the models themselves. The Anthropic SDKs (Python, Node, Go, Java, Ruby) are first-party and widely used — the Node SDK alone has ~5k weekly npm installs [2]. Prompt caching, batch processing, and the Files API are exposed through every SDK [3].

MCP — the Model Context Protocol — is the bigger structural bet. It's an open standard for connecting LLMs to tools, data, and other systems; the spec is on GitHub, the reference servers are TypeScript, and adoption has spread to OpenAI's Responses API, Google, and a long tail of indie tools [4]. For an AI lab, shipping a protocol other labs adopt is a different kind of distribution than shipping a model.

Internally, Anthropic uses Claude extensively to build Claude. Their own engineering blog posts describe Claude-driven code review, eval generation, and red-teaming in their training pipeline [5]. Less publicly: their Slack and internal documentation are reportedly Claude-augmented end-to-end. You can't verify this from outside; it shows up in offhand mentions in podcast interviews with the engineering team.

In the last six months, three things are worth tracking. First, the Claude 4.7 release in early 2026 added the 1M-token context window on Opus and a meaningful step in coding ability (their published SWE-bench and Aider numbers, with the usual caveats about benchmark contamination) [1].

Second, the Series F in early 2026 closed at a $61.5B valuation according to multiple outlets [2]. The notable thing isn't the headline number — it's that Anthropic kept Google and Amazon as strategic anchors while broadening the institutional base; both prior backers re-upped [2][3].

Third, MCP adoption crossed a real threshold: OpenAI's Responses API and Google's agent SDK both speak MCP as of Q1 2026 [4]. For a lab that didn't invent the model + agent framing, owning the connection layer between models and the rest of an engineer's stack is a quietly strong position.

Worth watching but not yet news: their hiring in London and Zurich suggests EU regulatory presence is being built; their public posture on the EU AI Act has been measured-but-engaged through 2025-2026.