Agents get an office: agentes-chat, numbrone and SuperARIS organise the team while blackhole and lasik keep it in a box

The repo calls itself an “Open-source, self-hosted Slack alternative where your team and AI agents work together in channels and DMs,” with agents running on Claude, ChatGPT/Codex, Gemini or Grok, MCP tools, their own computer and browser, routines, memory and approvals. It deploys with Docker Compose and Postgres and is AGPL-3.0.

Why it matters now: the interface for agents is converging on the one people already use for coworkers, a channel with threads and approvals. If agents are teammates, the chat log becomes the audit trail. At 1,159 stars in four days it is the biggest new repo in this window by attention, and attention is not proof it works. We did not run it.

Per the description: “A chief of staff for your coding agents: route requests to per-project manager agents that run parallel workers in git worktrees, on herdr.” Topics list claude-code, agent-skills and git-worktree.

Why it matters now: a three-tier shape, router, per-project manager, parallel workers, is how people are coping once one agent is not enough. The per-project manager is the useful idea, because it keeps context from bleeding between repos. The open question is who reviews the workers' output. We read the metadata only.

The repo describes “Your personal research agent with ARIS (Auto Research In Sleep), powered by a team of agents.” Topics name Claude Code, Codex, Gemini, MCP and herdr, so it targets several agent CLIs rather than one.

Why it matters now: overnight research is the most natural place for unattended agents, since a bad result costs a morning, not a deploy. Note numbrone above also builds on herdr, a sign a shared substrate for running agent sessions is forming. Early and small; the claims are the authors' own.

The description: “Skills for human-led AI teams: delegate tasks, verify evidence, explain steps clearly, and learn from real cases,” in English and Chinese. Topics include agent-skills, human-in-the-loop and workflow.

Why it matters now: after tooling for orchestration, this is tooling for the human manager. “Verify evidence” as a named skill acknowledges that delegating is cheap and checking is where the cost sits. It is a skills pack, so judge it by whether its prompts hold up in your workflow. We did not test it.

The one-line description is “A blackhole for your Claude Code.” Topics point to colima, container-isolation, virtio-fs, macos and sandbox, so the approach appears to be running the agent inside a container on a Colima VM.

Why it matters now: as agents run longer and with more permissions, the question shifts from “will it make a mistake” to “how far can the mistake reach.” A container with a mounted project is a cheap answer. The README is the place to check what is exposed; we read only the metadata.

Per the description: “BLA blinds your agent. LASIK gives it its sight back. Static de-blinding CLI + MCP server that undoes BLA (Blind LLM Agents) hardening so coding agents and REA can read the real code again. Never executes the target.” Topics include deobfuscation, prompt-injection and reverse-engineering.

Why it matters now: it implies a new arms race, code deliberately shaped to defeat AI readers and a counter-tool to read it anyway. Whichever side you are on, agents are now a threat model for code authors. The static-only, never-executes design is the right default for malware analysis. We have not verified the effectiveness claims.

The description: “Tell agents, AI crawlers, search crawlers and bots apart from people.” Topics include gptbot, user-agent, web-bot-auth and analytics.

Why it matters now: docs and product sites are increasingly read by agents first, and traffic reports that lump them with humans mislead. Classifying agent visits is the first step to serving them differently. It comes from a docs vendor with a stake in the answer, so read the rules before trusting the counts.

The repo’s description: “Use Claude Code as the coding agent in Xcode 27, powered by your own API key or gateway.” Topics include agent-client-protocol, acp and mcp.

Why it matters now: the topic tags suggest it speaks the Agent Client Protocol, which would let editors host third-party agents instead of one built-in assistant. If Xcode accepts that, bring-your-own-agent becomes the norm in IDEs. We did not confirm the Xcode 27 integration details.

Per the description: “A background virtual screen for any Android AI agent: it operates apps on an invisible display while you keep using your phone. CLI / MCP / HTTP, Shizuku, no root.” Topics include phone-agent, mcp-server and virtual-display.

Why it matters now: phone agents that grab your screen are unusable; ones that work on a virtual display are not. Exposing it over MCP makes it a tool any agent can call. Shizuku-level access is powerful, so treat the permission model as the thing to audit.

The description (in Portuguese): an open-source telephone exchange operated by AI agents that answers calls, runs tools during the call, consults a human and keeps working after the call ends; self-hosted. Topics include asterisk, sip, gemini-live and openai-realtime.

Why it matters now: voice is where agents meet customers with the least tolerance for error, so “consults a human” is the feature to look at. A self-hosted PBX plus realtime models lowers the entry cost for small operators. We read the description only and did not test call quality.

The description: “An Open-Source Personal Agent That Runs Fully on Your PC,” aiming at “efficient model–harness co-design” so everyday computers can run local AI. Topics include local-llm, proactive-agent and memory. The license field reads NOASSERTION, so check terms before use.

Why it matters now: it bets that the model and the agent harness should be designed together for small local models rather than a big cloud one. That is a claim to test on your hardware, not a result. Proactive behavior on a local machine also raises permission questions worth reading up on first.

The repo says it is “building the world’s most cutting-edge, CJK-first memory module for all mainstream agents.” Topics list bm25, sqlite, local-first, mcp and no-llm, so retrieval appears to be keyword-based on SQLite.

Why it matters now: most agent memory tools assume English tokenisation; Chinese, Japanese and Korean need their own handling. A local BM25-on-SQLite design is cheap, inspectable and offline. “Most cutting-edge” is the authors’ framing, and we ran no benchmark.

The description: “a Claude skill that turns P&ID PDFs (digital or scanned) into a process knowledge graph of flow, instruments and control loops.” Topics include isa-5-1, chemical-engineering and process-engineering.

Why it matters now: this is a skill as a vertical product, domain knowledge from one engineering field packaged for an agent. Piping and instrumentation diagrams are dense, so extraction accuracy is the thing to verify before relying on it. We did not test it.

Per the repo: “A Claude skill and TypeScript library for animated line-art SVG components: ink figures that walk and react, scenes with a day/night cycle.”

Why it matters now: the pairing of a skill with a real library means the agent generates against a stable API instead of improvising SVG each time. That is a pattern worth copying for any design system. At 583 stars it drew attention quickly; we did not view the output.