The agent got its own computer yesterday; today's tools are the shared rooms.

Block's self-hostable team workspace, open-sourced under Apache-2.0 and written in Rust: under the hood it is a Nostr relay where every message, reaction, workflow step, review approval, and git event is a signed event in one log — same shape, same identity model, same audit trail whether the author is a person or a process. The load-bearing design choice is that agents are members, not integrations: they get their own keypairs, their own channel memberships, and their own audit trail — scoped by identity, the way you scope a teammate, not by permission flags. Inside, agents can open repos, send patches, review code, run workflows, edit canvases, and drop into voice huddles; a feature branch can become a room where the patches, CI, review, and merge decision live together. Full disclosure: we picked Buzz on June 23, at 138 stars, as the collaboration layer of the agent-native stack — today it sits at 5,152 after a plus-3,252 day, with Dorsey's announcement two days old, which makes this the rare pick we get to re-run as a confirmation rather than a bet. On the 07-15 thread — the session becoming a room other people join — this is the inversion at team scale: the room is the unit of work, and both species are on the roster. Reach for it when your agents' output is scattered across terminals, bot DMs, and CI logs and you want the conversation, the patch, the workflow run, and the approval to be one searchable record. Delete the Slack-plus-webhook duct tape where the agent is a second-class citizen posting through a bot token. Tradeoff: what ships today is the single-relay setup, and adopting it is a workspace migration, not a tool install — Slack's gravity is real, and Nostr's identity model means key management for a whole team plus a fleet of agents is now your problem.

A macOS browser built for parallel occupancy: your agents run browser tasks in their own Spaces while your tabs stay yours, and the agent reaches your real logins, cookies, and extensions instead of starting from an anonymous profile. The authors name the incumbent contrast themselves — browser-use and agent-browser are automation frameworks that drive a separate browser, logins never carry cleanly, and you end up fighting the agent for tab focus; this is one browser designed for two occupants. Install is a DMG or `npx skills add citrolabs/ego-lite`, which drops an `ego-browser` skill into every agent's skills directory on your machine — the vendor-skill onboarding pattern from 07-16/07-17, now shipping from a browser. On first launch it offers to migrate your Chrome data wholesale. 1,246 stars since April, plus 219 today. Reach for it when browser agents keep dying on login walls, or when watching an agent commandeer your actual Chrome has lost its charm. Delete the second Chrome profile driven by browser-use and the re-authentication ritual that comes with it. Tradeoff: the MIT license covers the skill, the spec, and the docs — the browser itself is a closed binary from their CDN, so read "open source" narrowly; it is macOS-only; and handing an agent your real logged-in sessions is the 07-15 memory-heist lesson in browser form — the shared room's keys are your keys.

A slide tool that ships as a single ~560KB HTML file containing the editor, the viewer, the data, and the collaboration layer — MIT-licensed, free, no server, no account, works offline. The architecture is the pick: a JSON block near the top of the file holds the slide content where a program (or an agent) can read and edit it directly, and the app itself sits below, base64-compressed. Saves are ECDSA-signed; live co-editing runs through an encrypted relay on Cloudflare Durable Objects that cannot read the content. The origin story is today's theme in miniature — the makers were generating decks with coding harnesses and got tired of re-prompting an agent to nudge a text box, so they built the deck both species can operate: there is a Claude skill for generating Bento decks from scratch or converting PPTX. 881 points and 199 comments, the day's biggest Show HN. On the one-artifact-both-species thread, Nobie was the spreadsheet (07-14) and Tiptap the tracked-changes document (07-15); this is the deck, with the machine-legible second interface (07-10) literally at the top of the file. Reach for it when the agent drafts the deck and a human presents it — and every small edit currently means another prompt round-trip or a PPTX export. Delete the Google-Slides-account requirement for sharing a deck; recipients need a browser, nothing else. Tradeoff: the HN thread did the QA — multiplayer editing loses input focus, Firefox animations are choppy, images have no alt text, and there is a cloudflareinsights beacon on the homepage despite the privacy pitch. Single-file is the product and also its ceiling.

A hosted gateway that gives agents scoped access to your live databases — Postgres, MySQL, ClickHouse, and PostHog — with row-level and table-level restrictions enforced per role, over MCP, without ever exposing credentials to the agent. Query results can become HTML dashboards, and the detail that earns the pick is that viewer permissions are re-checked at load time: a dashboard built by someone with broad access shows a narrower viewer only what their role allows. The maker's operational note is its own small data point in this week's cost arithmetic: query execution runs on a cheap-but-competent model because tool-calling at scale is a unit-economics decision now. 70 votes and #31 on Product Hunt today. On the enforcement-beats-instruction thread (deptrust 07-03, MakerChecker 07-07, Kastra 07-10), this is the pattern arriving at the database layer — and it rhymes with Buzz upstairs: scope by identity, not by hoping the prompt holds. Reach for it when the agent needs production data and your current answer is a read replica and a prayer. Delete the read-only service-account credentials pasted into an MCP config — the ones that are read-only on every table, including the ones the agent has no business reading. Tradeoff: hosted and closed — the walls are real but the building is theirs, and routing your production query path through someone else's platform is exactly the dependency yesterday's self-hosted slate was escaping.

An MIT-licensed web terminal multiplexer for agent sessions: a kanban board of every active agent, push notifications when one finishes or blocks on input, a built-in file browser and editor, git worktrees for parallel work, and a voice mode — across OpenCode, Claude Code, Codex, Copilot, Antigravity, and Pi. 71 votes and #30 on Product Hunt; the repo is three weeks old at 34 stars. This is a watch closing at maximum speed: yesterday's pick was AgentManager, the $48-a-year, Claude-Code-only, single-machine doorbell, and yesterday's note said a free open-source clone of the pattern was one motivated weekend away. Caw landed the cross-machine, cross-harness, web-based version the next day, free — and it completes today's slate the same way AgentManager completed yesterday's: four shared rooms above, and this is how the rooms reach you after you walk out of them. Reach for it when agents run on more than one machine under more than one harness and you still want to see the whole wall from your phone. Delete the cmd-tab patrol, and reconsider the AgentManager subscription if cross-harness coverage matters more to you than notch-level polish. Tradeoff: three weeks old, 34 stars, and a web surface over every terminal your agents hold — sessions are full of secrets (07-15), so however you deploy it, the auth in front of it is load-bearing.

Off the thread but worth knowing: **Chimlo** (GPL-3.0, free, no telemetry by default) is the other half of the doorbell watch, landed the same day as Caw — a native macOS app that puts Codex and Claude Code session status in the MacBook notch and lets you approve Claude Code permission requests from it; 63 votes on Product Hunt, repo three days old. AgentManager's floating window, free and open, three days after we priced the weekend. **Millwright** (Apache-2.0, one Rust binary, v0.1.0) and **Cactus Hybrid** (MIT, three days old, 147 points on the front page) answered who-gets-the-call twice in one day from opposite directions: Millwright is a self-hosted LLM router with cheap/mid/frontier roles, per-session provider affinity to keep prompt caches warm, and no hosted control plane — pitched, in the maker's own Show HN text, against the hosted routers in the week OpenRouter is rumored to be acquired; Cactus Hybrid moves the router inside the checkpoint, post-training Gemma 4 E2B to emit a structured confidence score so 15–35 percent of queries hand off to a frontier model and the rest stay on-device — the handoff table is the author's own, with quantization degradation stated in the open. **OpenCodeReview** (Apache-2.0, 11k stars since May, first crossed our pool June 5) is Alibaba open-sourcing its internal AI code-review CLI after two years and tens of thousands of internal developers — agentic line-level review that works with Claude Code, Codex, and Cursor, plus an `ocr scan` mode for auditing code with no diff: the machine-verification thread (07-17, 07-20) at its largest vendor scale yet. And **openship** (Apache-2.0, 7.7k stars, plus 1,302 today) is a self-hosted deployment platform whose control plane is a desktop app driving your servers over SSH — trending the same day as Instatic (self-hosted CMS) and Megaphone (open-source on-device Mac dictation): the self-hosted current running under today's shared-rooms slate.