# @anthropic-ai/sdk Projects at AI Tinkerers

> Canonical HTML: https://aitinkerers.org/technologies/anthropic-ai-sdk
> Markdown URL: https://aitinkerers.org/technologies/anthropic-ai-sdk.md
> Technology record last updated: 2026-04-08T19:25:48Z
> Generated: 2026-09-20T19:47:24Z

Official TypeScript client for Claude: deploy models like Claude 3.5 Sonnet into Node.js or browser environments with native type safety.

This SDK serves as the primary gateway for Anthropic's Claude models (3.5 Sonnet, 3.0 Opus) across Node.js, Bun, and browser runtimes. It handles the heavy lifting for streaming responses via Server-Sent Events (SSE), processing vision inputs, and executing tool use (function calling) for agentic workflows. You get production-ready features out of the box: automatic retries, custom timeouts, and strict TypeScript definitions for every API endpoint. Get started with a single command: npm install @anthropic-ai/sdk.

- Official technology site: https://github.com/anthropics/anthropic-sdk-typescript
- Public AI Tinkerers demos and talks: 2
- Result page: 1 of 1

## Recent Public Talks and Demos

### [Your BD Pipeline Is A Good First Autonomous Agent](https://boston.aitinkerers.org/talks/rsvp_Vjy6qYY7UX0)

An end-to-end BD pipeline that runs itself on a weekly cron: it sources leads, researches them, drafts the outreach, and emails me the two or three things to do this week. Sourcing → research → draft → nudge, with a lightweight CRM to run it from. One rule holds everywhere: the system researches and drafts; a human always sends. Why this is worth five minutes of a GTM room’s time: almost every agent demo shows one primitive at a time — here’s tool use, here’s a cron, here’s memory. That leaves you knowing the parts and not the shape. BD is a workflow everyone in this room already has opinions about, which makes it a good place to see the whole surface area doing real work at once. The demo is a guided tour of that: I trace one target end to end, and name the primitive earning its keep at each step. The tour, following one company through the system: It gets found. The Opportunity Radar treats my ICP as a graph rather than a list. It starts from private-equity sponsors active in my space and walks outward — a locked sub-agent maps each fund’s portfolio in parallel threads, and qualified portfolio companies auto-feed the pipeline. The frontier of unvisited sponsors lives in a memory store, so the walk resumes each Monday instead of cold-starting. Ownership structure is what makes a buyer resemble my last buyer, so the graph edge qualifies better than a keyword match. It gets carded. The Radar pushes the target into my CRM over MCP — the CRM is itself an MCP server, so the agents and I drive the same pipeline through the same tools. It gets researched. A graded session produces a seven-section pre-call brief off the live web. The run ships with a rubric; the platform scores the deliverable and sends the agent back to revise before I see it. The verdict lands on the card. It gets formatted. Skills load the output contract on demand — including a bundled validate_blocks.py the agent runs against its own draft so the CRM can parse the people and portfolio blocks. It gets drafted. A nightly cron drafts outreach from a compiled brand pack. Secrets come from a vault injected at egress — the agents never see a key. It stops. The draft sits at the gate. Nothing sends. Pipeline status is never machine-written. What I’ll show live: the Monday-morning board with fit scores and grader verdicts that arrived without me; one target’s drawer opened — why-now, signals, key people, dossier, draft; the sweep receipt showing what the Radar covered and where it stopped; a rubric and the revision it forced; and the eval scorecard where a prompt change shows up as a delta against a pinned baseline.

- Event context: Back from Summer: AI GTM Builders — 2026-09-03 — Boston
- Public talk page: https://boston.aitinkerers.org/talks/rsvp_Vjy6qYY7UX0

### [What did we learn from the March 31st Claude Code Source Leak?](https://seattle.aitinkerers.org/talks/rsvp_P82HqL1GQaM)

This technical analysis, presented by Daniel Motles of Qumulo in April 2026, dissects the architecture and leaked details of Claude Code. Claude Code is a Bun-native TypeScript application with a React and Ink terminal UI, functioning as a productized agent operating system. The system is composed of 1,902 source files totaling 512,685 lines of code. Its technical stack includes the Anthropic SDK, MCP integrations, and native tools for Git, Shell, and LSP. A major finding is the discovery of KAIROS, an always-on autonomous daemon that runs heartbeat prompts every 30 seconds to identify tasks, fix errors, and manage memory consolidation. Other unreleased features include Ultraplan for deep thinking via remote Opus variants, Coordinator Mode for spawning sub-agent forks, and an internal Undercover Mode that strips AI signals from public commit histories. The system utilizes a heavyweight persona driven by a massive system prompt and a four-tier MD memory system covering users, feedback, projects, and references. To preserve context window space, Claude Code uses deferred tooling and a local LRU cache to prevent redundant token consumption. Local search is handled by a parallelized Rust-based engine called RipGrep (rg). The analysis concludes that the engineering "plumbing"—the tools, memory, and safeguards—constitutes the product's primary competitive moat. While the leak has spurred open-source clones, it has also raised enterprise concerns regarding release hygiene and intellectual property safety.

- Event context: AI Dev Tool Track — Seattle Meetup - April 13th, 2026 — 2026-04-14 — Seattle
- Public talk page: https://seattle.aitinkerers.org/talks/rsvp_P82HqL1GQaM

## Related Technologies

- [React](https://aitinkerers.org/technologies/react) ([Markdown](https://aitinkerers.org/technologies/react.md)) — 219 public demos
- [Bun](https://aitinkerers.org/technologies/bun) ([Markdown](https://aitinkerers.org/technologies/bun.md)) — 19 public demos
- [Claude Managed Agents](https://aitinkerers.org/technologies/claude-managed-agents) ([Markdown](https://aitinkerers.org/technologies/claude-managed-agents.md)) — 2 public demos
- [Claude Opus](https://aitinkerers.org/technologies/claude-opus) ([Markdown](https://aitinkerers.org/technologies/claude-opus.md)) — 18 public demos
- [Commander](https://aitinkerers.org/technologies/commander) ([Markdown](https://aitinkerers.org/technologies/commander.md)) — 4 public demos
- [Next](https://aitinkerers.org/technologies/next) ([Markdown](https://aitinkerers.org/technologies/next.md)) — 186 public demos
- [TypeScript](https://aitinkerers.org/technologies/typescript) ([Markdown](https://aitinkerers.org/technologies/typescript.md)) — 205 public demos