😸 Agents go postal

gm legends, happy Thursday.

Today: the open-source workspace trying to swallow your whole tool stack, one API that turns the entire internet into agent food, and an MCP server that lets your AI send actual paper mail. Plus the forum on what breaks first when a vibe-coded project meets reality.

Macro is Jacob Beckerman's answer to the stack that ate his last startup, Slack plus Linear plus Notion plus 17 other tools: an open-source workspace that pulls email, docs, messaging, tasks, and video calls into one app with a single AI memory across all of it, so the context from a call transcript, an email thread, and a doc all live in the same brain.

🔥 Our Take: Every tool your team uses now has its own AI, and each one has read exactly one folder. Nine assistants, none of them knows what the other eight saw. That's the thing Macro is calling dumb: merge email, docs, chat, tasks, and calls into one app, and you get one AI that's seen all of it, the email thread, the call transcript, the task nobody picked up. And Beckerman isn't a kid with a landing page, he spent years teaching software to read legal contracts, built the Macro PDF editor, and raised $9.3M from a16z on it. Notion is circling the same bundle from the docs side with AI sprinkled on top. It all lives or dies on whether a real team will rip out five tools it already pays for, and I've never watched one manage it.

Context.dev is one API for the messy job every AI team keeps rebuilding: scrape any URL, crawl a site, get clean LLM-ready markdown, pull structured data, grab screenshots and logos, enrich companies. Yahia Bakour's YC-backed team already powers Mintlify, daily.dev, and Chatwoot, and an agent can even sign itself up and wire in the API on its own.

🔥 Our Take: Scraping is the job every AI team swears is two days of work and then maintains forever. Rotating proxies, broken selectors, markdown cleanup, it never ends. Context.dev sells the whole mess as one API, and the customer list backs it up, Mintlify and daily.dev run on it, and 90% of signups arrive having just torched their homegrown scraper. Best detail: your coding agent can sign itself up, grab a key, and wire it in without you, which tells you exactly who they think the customer is in two years. It's a brutal market, Firecrawl's right there and half of YC besides, and the whole category depends on websites not winning the bot-blocking war. But nobody has ever loved maintaining a scraper.

ElevenLabs has been the go-to for voice for a while. Now they've turned that expertise into agents that actually get things done. You set one up, it talks like a real person, listens, responds, and helps handle the task — support calls, bookings, whatever the job is. Not a demo, not a "press 1 for sales" situation. It's ready to deploy. Feels like one of those shifts where the interface quietly changes. Less typing, less clicking, more just saying what needs to happen and letting it play out.

PieterPost MCP connects your AI agent to the actual postal system: it can quote a letter or postcard, upload the file, spin up a checkout link, and a real piece of paper lands in a real postbox. It's an MCP server from PieterPost, the Dutch send-letters-online service named after the Netherlands' Postman Pat.

🔥 Our Take: Your agent can send actual mail now. Paper, stamp, letterbox, via PieterPost, the Dutch letters service named after their Postman Pat. It's funny until you notice it's the hardest trust case in agents, because mail can't be unsent, a wrong address isn't a retry, it's a stranger reading your rent dispute. Their setup gets it right: the agent drafts, quotes, and preps the order, and nothing prints until a human pays. Same shape as Claude ordering DoorDash last week, the agent does the legwork, you press the scary button. The quiet story is a tiny Dutch company just walked itself into every Claude user's toolbox with one MCP server, no app store, no BD deal, and small companies are going to figure that trick out long before big ones do. This one's just a joy.

Omri Ben-Shoham (@omri_ben_shoham1) asked the question every vibe coder eventually lives: what broke first when your AI-built project met reality? The answers read like a group therapy session with receipts.

The failures split into camps. Architecture rot, duplicated logic and random utility functions the AI keeps patching instead of redesigning. Databases where the model adds a new table instead of simplifying the old ones. And money: Erika (@erika_chen) burned $170 in API usage on what should have been a simple reporting script, and still doesn't know exactly where the tokens went.

The best story came from @redist, whose voice agent kept cheerfully telling callers "you're all booked!" while the calendar tool was silently disconnected, so nothing was ever booked. Nothing in the transcript looked wrong. The lesson the whole thread rallied around: never trust the summary, check the system of record. The agent's story about what happened and what actually happened can diverge silently, and confidently, at the same time.