Open-source Claude Design

Today’s top AI Highlights:

1. Open-source replica of Claude Design

2. Cursor runtime, harness, and models now in Cursor SDK

3. Give your agent a filesystem, it gets a memory

4. Warp is now open-source

5. Convert any document into MD or JSON for agents

& so much more!

Read time: 3 mins

AI Tutorial

Anatomy of Agent SKILLS

Your agent has a 200k token context window.

The 400 tokens of instructions it actually needs are buried under tool definitions, reference docs, and brand guides it never asked for. So it ignores them.

This is the most common reason agents fail in production. It's not a model problem or a framework problem.

In this blog, you'll learn the anatomy of Agent Skills: why the first two lines of SKILL.md are the most important writing you'll do, and how the LLM itself routes queries to the right skill without embeddings or retrieval layers.

You'll also see how skills enable independent team ownership and composition at scale, doing for agents what npm did for JavaScript.

Read on to learn the five parts that make skills work, then pick one workflow you do every week and ship your first skill today.

We share hands-on tutorials like this every week, designed to help you stay ahead in the world of AI. If you're serious about leveling up your AI skills and staying ahead of the curve, subscribe now and be the first to access our latest tutorials.

Don’t forget to share this newsletter on your social channels and tag Unwind AI (X, LinkedIn, Threads) to support us!

Latest Developments

Open-source Replica of Claude Design

When something good ships closed-source, the clock starts on the open-source version.

Open Design is a local-first replica of Claude Design that ditches the lock-in entirely.

The trick is that it doesn't ship an agent at all. The daemon scans your machine for Claude Code, Codex, Cursor Agent, Gemini CLI, OpenCode, or whichever it finds. That same agent gets driven by a stack of 19 skills, 71 brand-grade design systems, and a sandboxed iframe preview.

Same artifact-first loop as Claude Design. None of the cloud, subscription, or model lock-in.

Key Highlights:

1. BYO agent - Whatever CLI is on your PATH becomes the design engine. No Anthropic key needed if you have Claude Code installed; no Claude Code needed if you have Codex or Gemini.

2. 19 Skills + 71 design systems - Prototypes, decks, dashboards, mobile apps, plus brand systems for Linear, Stripe, Vercel, Notion, Apple, Tesla, and 65 others.

3. Anti-slop, hard-coded - Turn 1 is always a question form. Turn 2 picks one of five locked visual directions. The agent self-critiques 1-5 before emitting.

4. Skills are just folders - Drop a SKILL.md into skills/, restart, it appears in the picker. Same convention as Claude Code.

Cursor Runtime, Harness, and Models now in Cursor SDK

The agents that power your Cursor desktop app, CLI, and web app are now just a npm install away.

Cursor has launched the Cursor SDK in public beta, giving you programmatic access to the same runtime, harness, and models that ship inside Cursor itself. Build custom coding agents in a few lines of TypeScript, run them locally or on Cursor's cloud against a dedicated VM, and pick any frontier model you want.

You get codebase indexing, semantic search, MCP server support, skills auto-loaded from .cursor/skills/, hooks for controlling the agent loop, and subagents that delegate tasks with their own prompts and models.

Cursor has also open-sourced starter projects, including a coding agent CLI, a prototyping tool, and an agent-powered kanban board.

Give Your Agent a Filesystem, It Gets a Memory

“RAG is dead” and “filesystems are all you need” are both kind of right and kind of wrong.

Agentic search on a codebase is genuinely beautiful - until you try it on folders with your notes, PDFs, transcripts, etc., and watch it fail.

SMFS is what happens when you stop picking sides. From Supermemory, it’s a mountable filesystem that turns your memory container into a real local directory. Your agent navigates it with the same ls, cat, and grep it already knows, except grep is now semantic by default.

Drop in PDFs, videos, audio, and docs - SMFS handles extraction, embeddings, and search automatically.

It’s open source, written in Rust, and plays well with Daytona, E2B, Cloudflare, and Vercel sandboxes.

Key Highlights:

1. Semantic by default - Inside a mount, plain grep hits Supermemory’s hybrid semantic index; any flag falls through to standard grep.

2. Multi-agent, shared state - Multiple agents can mount the same container - one writes, the others see it on the next pull.

3. Live profile file - A virtual profile.md at the mount root regenerates on read from every memory in the container, so the agent’s first useful action in a new directory - reading a file - just works.

4. Works where there’s no filesystem - For Cloudflare Workers, edge runtimes, and browser agents, @supermemory/bash exposes the same Unix surface as a single run_bash tool the agent calls.

Quick Bites

Warp is now open-source
Warp is now open source, dropping its client code, roadmap, and a lightweight contributor workflow on GitHub under a mix of MIT (UI crates) and AGPL v3 (everything else). OpenAI is the founding sponsor, GPT models drive the new agent management flows, and Warp still plays nicely with Claude Code, Codex, and Gemini CLI if you'd rather BYO agent. The cherry on top: build.warp.dev lets you spectate fleets of Oz agents triaging issues and shipping PRs, then jump into a live web-compiled Warp terminal to follow along.

A2A + MCP: Five Ways They Fit Together
This one is a solid playbook on how A2A and MCP fit together. Five multi-agent patterns covering everything from agent discovery to ambient event meshes that react to Pub/Sub streams in the background. The TL;DR: A2A handles agent-to-agent, MCP handles agent-to-tool, and if you're not using both you're going to be rewriting things in twelve months.

The five patterns:

1. Agent Card Discovery — every A2A agent publishes a JSON doc at a well-known URL describing its capabilities, auth, and rate limits; Agent Registry makes them findable across your org

2. Delegated Specialization — a Python coordinator hands off to a Go security agent, a Java risk agent, a TypeScript marketing agent; each team ships independently and the coordinator never knows the difference

3. Tool Bridge via MCP — one protocol for every data source and API, with 60+ ready integrations and Apigee API Hub auto-converting existing REST APIs into agent-accessible tools

4. Cross-Organization Federation — Agent Gallery's 100+ validated partner agents (Adobe, ServiceNow, Workday, Salesforce) collaborate with yours while each side enforces its own governance

5. Ambient Event Mesh — agents listen continuously on BigQuery tables and Pub/Sub streams, processing events in the background and escalating to humans via Mission Control when needed

Worth a read if you're stitching agents together right now and tired of building custom connectors for every API.

Open-source spec for Codex orchestration
OpenAI just open-sourced Symphony, a spec that turns your Linear board into a control plane for Codex agents — every open ticket gets its own agent workspace, runs continuously, and shows up as a PR for review. Internally, some teams saw landed PRs jump 500% in three weeks, though the more interesting bit is the mental shift: engineers stop babysitting sessions and start managing work. It's shipped as a Spec.md you point at your coding agent of choice, with an Elixir reference implementation — fitting, given the whole thing was built by Codex itself.

Tools of the Trade

1. Firecrawl /parse: New endpoint that converts documents like PDFs, DOCX, and XLSX into clean Markdown or JSON that AI agents can actually work with. It runs on a Rust-based engine that's up to 5x faster than before, handles files up to 50 MB, and preserves reading order and tables.

2. Gemma Chat: Vibe code without internet. This open-source Electron app runs Gemma 4 natively on Apple Silicon. You describe what you want to build, and it writes the code with a live preview that updates as the model types. No internet connection needed after the initial model download.

3. /browser-trace skill: Attaches a read-only observer to your agent's browser session, capturing network requests, DOM snapshots, screenshots, and CDP logs into a filesystem. It pairs with Stagehand, Playwright, or whatever's already driving the page so you can debug failed runs, attach mid-flight to live automations, or just keep tabs on what your agent is actually doing.

4. Helmor: A local-first macOS IDE built around orchestrating coding agents rather than writing code yourself. Built for agent orchestration, review, testing, merge, and everything around the code.

5. Awesome LLM Apps - A curated collection of LLM apps with RAG, AI Agents, multi-agent teams, MCP, voice agents, and more. The apps use models from OpenAI, Anthropic, Google, and open-source models like DeepSeek, Qwen, and Llama that you can run locally on your computer.
   (Now accepting GitHub sponsorships)

That’s all for today! See you tomorrow with more such AI-filled content.

Don’t forget to share this newsletter on your social channels and tag Unwind AI to support us!

Unwind AI - X | LinkedIn | Threads

Awesome LLM Apps | Sponsor Us

PS: We curate this AI newsletter every day for FREE, your support is what keeps us going. If you find value in what you read, share it with at least one, two (or 20) of your friends 😉