Postman Collections in Google Antigravity IDE MCP Store

Cloudflare is rebuilding Wrangler into a unified CLI that exposes all ~3,000 API operations across 100+ products through a generated, consistent interface designed primarily for AI agents. Local Explorer adds a local dashboard to introspect simulated resources (KV, R2, D1, Durable Objects) during development, using the same API structure as production. Together, these tools address the shift toward agents as primary API consumers by eliminating inconsistencies that cause agents to fail and by giving developers visibility into what their agents are doing locally. Discover the benefits of the CF CLI Local Explorer tool in our recent blog article. Take your Cloudflare experience to the next level with this handy command-line interface feature.

𝗪𝗵𝗮𝘁 𝗶𝗳 𝘆𝗼𝘂𝗿 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁 𝗰𝗼𝘂𝗹𝗱 𝗮𝗰𝗰𝗲𝘀𝘀 𝟮,𝟱𝟬𝟬+ 𝗔𝗣𝗜𝘀... 𝘂𝘀𝗶𝗻𝗴 𝗷𝘂𝘀𝘁 𝟭,𝟬𝟬𝟬 𝘁𝗼𝗸𝗲𝗻𝘀? That’s exactly what Cloudflare is enabling with its new Code Mode MCP server. 𝗧𝗵𝗲 𝗽𝗿𝗼𝗯𝗹𝗲𝗺: Traditional MCP setups expose every API endpoint as a tool, quickly exploding context size and burning tokens before the agent even starts reasoning. 𝗧𝗵𝗲 𝘀𝗵𝗶𝗳𝘁: Cloudflare replaced thousands of tool definitions with just two primitives: • search() to discover APIs • execute() to run generated code Instead of loading schemas into context, the agent writes code against a typed SDK. 𝗧𝗵𝗲 𝗿𝗲𝘀𝘂𝗹𝘁: • Token usage drops ~99.9% (from ~1.17M → ~1K) • Context size becomes constant, not proportional to API surface • Agents can orchestrate multi-step workflows in a single execution cycle 𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀: This flips the model from tool calling → code generation as orchestration Less context overhead. More reasoning space. Real agent workflows. One line that captures it: “𝘊𝘰𝘯𝘷𝘦𝘳𝘵 𝘵𝘰𝘰𝘭𝘴 𝘪𝘯𝘵𝘰 𝘤𝘰𝘥𝘦 𝘢𝘯𝘥 𝘭𝘦𝘵 𝘵𝘩𝘦 𝘮𝘰𝘥𝘦𝘭 𝘱𝘳𝘰𝘨𝘳𝘢𝘮 𝘪𝘵𝘴 𝘸𝘢𝘺 𝘵𝘩𝘳𝘰𝘶𝘨𝘩 𝘵𝘩𝘦 𝘈𝘗𝘐.” Are we moving from “LLMs calling APIs” to “LLMs writing programs over APIs”? For more details, read InfoQ news article: https://lnkd.in/gicGMc3M

This flips agent design from brittle tool wiring to programmable execution. Slashing context from millions of tokens to ~1K removes a hard scaling ceiling. “search + execute” mirrors how engineers actually navigate large API surfaces. This is a strong step toward agents that behave like systems, not wrappers.

When I started to build Squirrel Notes, I wanted to explore building software with AI but also look into creating an MCP server, it's actually turned into a key part of my workflow. For the past few weeks, Squirrel Notes has been the persistent memory layer for my Claude workflows. Every week, Claude pulls my Google Calendar, structures a prep note - working locations, pending actions, day-by-day schedule, key context - and saves it directly to Squirrel. Same for time allocation reports, sprint health checks (with Jira integration), and meeting summaries. The thing that makes it work isn't the AI. It's having somewhere the output lands and stays. Claude has no memory between conversations. Squirrel is the memory. Reusable prompts live there. Programme notes live there. Weekly prep lives there. I'm not rebuilding context from scratch every time I open a new chat. How it works The connection runs over MCP stdio - a local process on your machine, not a hosted server. That matters because Squirrel is zero-knowledge. The server stores ciphertext. Your passphrase never leaves your browser. Decryption happens on your device. A hosted MCP server would undermine that entirely. With stdio it doesn't. The trust boundary is your local machine - the same boundary you already accept when you open the app. Config is a few lines in Claude Desktop. Claude gets tools to read, write, search, and organise notes. Every access generates an audit event. You can see exactly what was touched. https://lnkd.in/gxK-_p2B

For developers building the "Agentic Web," observability is the biggest hurdle. When an AI agent executes code in a remote sandbox, you need the same level of control you have on your local machine. Today we’re launching the Cloudflare CLI and Local Explorer. These tools provide a bridge between your local development environment and our global edge, allowing for real-time debugging, terminal access, and filesystem inspection of AI agent workloads. #AgentsWeek https://cfl.re/4tHEyB5

This is huge for Cloudflare! Our team at Vigilbase is already exploring how we can leverage the Cloudflare CLI via our Aegis agentic platform, to help customers manage their Cloudflare resources at light speed. Keep an eye out for news soon! 👀

Update: Pulsedive now has a dedicated documentation site with a complete API reference, live playgrounds, and MCP server support for AI-assisted development. Check it out: https://lnkd.in/eGX8JUm4 Announcement: https://lnkd.in/egDhks4t

Firecrawl Alternative: Same API, 5.5x Faster, Self-Hostable. The migration takes 5 minutes. Here's why. CRW implements a Firecrawl-compatible API. Same endpoints, same response format. You change one line — the base URL — and everything works. Before: const client = new FirecrawlApp({ apiUrl: "https://api.firecrawl.dev" }) After: const client = new FirecrawlApp({ apiUrl: "http://localhost:3002" }) That's it. No SDK changes. No schema migration. No rewrite. But the numbers change dramatically: Latency: - Firecrawl: 4,600ms avg - CRW: 833ms avg (5.5x faster) RAM: - Firecrawl: ~500MB - CRW: 6.6MB (75x less) Content coverage: - Firecrawl: 77.2% - CRW: 92% Docker image: - Firecrawl: 2GB+ - CRW: 8MB Cold start: - CRW: 85ms Plus: - Fully open source (AGPL-3.0) - Self-hostable on any infrastructure - Built-in MCP server for AI agent integration - No API keys, no rate limits, no vendor lock-in If you're hitting Firecrawl's rate limits, paying for usage you could self-host, or need lower latency — the switch is one line of code. github.com/us/crw | fastcrw.com #FirecrawlAlternative #WebScraping #OpenSource #SelfHosted #DevTools

Authentication is the hardest part of building MCP in production. Not because it's inherently complex, but because there are actually two completely separate problems that most guides conflate into one: 1. How does the AI client (ChatGPT, Claude, Cursor) authenticate to your MCP server? 2. How does your MCP server authenticate to your backend APIs? We just published a complete breakdown at Manufact (YC S25), based on real implementation. The patterns are very different depending on whether you're building a public MCP server, an internal enterprise tool, or something in between. We cover all three. If you're building MCP servers that connect to anything beyond a public API, this is the guide I wish existed when we started. https://lnkd.in/dGVrtwwK

A talk by Sunil Pi at Cloudflare has been on my mind this week. It's called Code Mode. Short version: stop doing JSON tool-calling with your LLM. Let the model write actual code. Run it in a sandbox next to your APIs. The proof point is wild. Cloudflare has 2,600 API endpoints. Traditional tool exposure = 1.2M tokens on the first call. Two code-executing tools = 1,000 tokens. 99.9% reduction. Eight round-trips become one execution. Here's what caught me. This is already how I work with Claude skills. When I find something repetitive that works, I ask Claude to write a script for it. Save it inside the skill. Reuse it forever. Claude only touches the script when the logic actually needs to evolve. That's how you cut token consumption without cutting capability. If you're building with Claude and you're not doing this yet, start. Every recurring pattern in your workflow is a script waiting to be generated once and reused a thousand times. The part of the talk that opened a new door for me: Generating code on demand for your customers. Picture onboarding a new B2B account. The AI builds them a custom interface shaped to their workflow, their data, their specific needs. Generated once at onboarding. Stored. Served for the lifetime of the account. That's a genuinely interesting idea. The caveat: generate once = economical. Regenerate on every session = token bill from hell. The principle I keep coming back to: Generated code is an asset, not exhaust.