-
Posts
-
Open-source Jev that runs on CPU
+ Grok Bot, Muse, and Instinct agents talk to each other #
Start here ↓
Get your Grok Bot, Muse, and Instinct agents to talk to each other.
Grok Bot is always on in the cloud. Muse can make phone calls. Codex and Claude Code know your code, while ChatGPT and Claude hold years of conversations.
Matt Van Horn built Agent Tincan, which lets agents in completely different products ask one another for help. This means Grok Bot can ask Muse to call a restaurant, any agent can retrieve an old Claude chat, and the reply travels back without you becoming the copy-paste layer.
To get started, you need a Tailscale network and one always-on Mac or Linux machine to host the relay. Set it up in less than 5 minutes.
🚀 Shipped #
Run this tiny open-source alternative to Jev on a CPU. Supersonic Labs released Julia 1, a 144.3M model that, just like Jev, takes context, a question, and a changing set of possible answers, then returns scores instead of generating prose. It is an independent model built on the multilingual mmBERT-small, not on Jev weights or a Qwen fine-tune. Runs entirely on CPU, including about 5 decisions per second on a Samsung tablet. Announcement blog | Hugging Face
TypeSafe AI and OpenRouter shipped typesafe/jev-router, a cache-aware router powered by Jev. Before an LLM call, Jev scores the request and picks the model plus reasoning effort, balancing quality, speed, and cost. The router can avoid unnecessary generation by making that decision first instead of spending tokens to discover which model should have been called.TypeSafe: Jev Router
Turn your REST APIs into MCP tools with Google Cloud API Gateway. Annotate the OpenAPI spec you already deploy, and the gateway serves your REST operations as MCP tools on the /mcp path without a separate server. One detail to catch before production: tools/list is unauthenticated by default, and Google says discovery must be secured with JWT because API keys cannot protect that method.Google Cloud announcement
Claude Code will now try to find a graceful stopping point when a five-hour session limit hits mid-task instead of cutting off halfway through an edit. It gets a small, fixed allowance from the weekly limit to wrap up what it can. The rollout gives Pro users one wrap-up a week, while Max and Team Premium users get one whenever they hit the five-hour limit.Claude Code announcement
GLM-5.3-Flash can imitate a Jev-style decision model. Privatemode showed how to turn GLM-5.3-Flash into a Jev-like decision system by prompting it so the first output token answers the question. That produces a typed decision in a single forward pass, but it is still a prompting technique applied to a standard LLM, not a purpose-trained decision model. So Jev remains several times cheaper.Privatemode blog
Ollaya is Ollama for local decision models. This open-source local runtime comes with a desktop app, CLI, Docker image, and TypeSafe-compatible /v1/systemone endpoint. It runs models such as Laya, NLI, GLiClass, Qwen3Guard, and Decider on CPU or NVIDIA GPU, with Apple GPU support for some models. Ollaya
Google released Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS for expressive speech generation. Flash handles custom voices, character design, and line-by-line performance direction, while Flash-Lite targets dubbing, voice agents, and bulk audio. Voice replication is restricted in several regions, including the EEA, UK, Switzerland, India, Texas, and Illinois.Google announcement
Google Antigravity agents can now run local models. Google added local-model support to the Antigravity SDK, starting with Gemma 4 26B A4B through LiteRT. You can now run agent workflows offline or combine a cloud planner with local workers. Google announcement
🧠 Worth Knowing #
DeepSeek published the sandbox layer behind agent training at scale. DSec paper describes the sandbox infrastructure it uses for large-scale agentic training. It exposes function-calling, container, microVM, and full-VM backends through one SDK, then handles placement, lifecycle, memory sharing, image , and reward-hacking mitigation. DeepSeek paper
Claude solved the Nine Loops physics problem experts thought would need far more compute. A physicist challenged AI labs to solve a calculation from his former field that no one had completed because it looked too computationally expensive. A month later, Claude Fable 5.1 produced the nine-loop result in two different ways, and physicist Lance Dixon independently validated it. Anthropic blog
Claude plugins now have a submission portal. Anthropic opened a developer portal for submitting Claude plugins to its directory. The portal handles validation, safety scans, review feedback, publishing, and post-launch usage analytics. Anthropic blog
Some details of how OpenAI agents hacked Hugging Face. Swarmtraces published this investigation into how OpenAI agents escaped restricted evaluation environments and reached HF systems. More than 80,000 redacted attack payloads were reconstructed from public link chains, covering service chaining, internal Slack searches, attempted evidence deletion, and exposure of sensitive data that Hugging Face says it revoked.Swarmtraces blog
🔧 Clone and Run #
Clone & Run of the Day
Generate and edit videos with Claude Code. video-use by Browser Use is an open-source project that lets Claude Code, Codex, Hermes, or OpenClaw turn raw takes into a finished MP4. It requires FFmpeg, video dependencies, and an ElevenLabs key for transcription.video-use
Run jevgrep when your coding agent is wasting tokens finding files. Coding agents spend part of every unfamiliar task finding the right files. Jevgrep uses Jev to find the relevant files. Just ask a repository question, and jg returns relevant files, reading leads, and verbatim source excerpts in one stdout response. Your coding agent then implements and tests the change.jevgrep
Clone Takeone so your agent records the demo for you. Takeone is like Screen Studio for agents. Use the skill and CLI for recording web-app demos with smooth cursors, zooms, click ripples, and headless rendering. Takeone
Run Tobi Lütke’s disk treemap when your storage vanishes. disktree helps you find and remove what fills a disk. It scans your home directory by default, draws every directory as a nested mosaic sized by what it really costs on disk, and lets you walk into it with the keyboard or the mouse.disktree
Clone a bug tracker that lives inside Git. git-bug is a bug tracker that stores issues, users, and comments as Git objects. It includes CLI, terminal, and web interfaces plus bridges to GitHub and GitLab.git-bug
Run a Slack teammate that remembers the company. Company Brain is Supermemory’s open-source Slack agent that remembers conversations and can work through GitHub, Linear, Notion, Google Workspace, MCP, and a sandbox. It deploys to your Cloudflare account and requires Workers and D1 setup.Company Brain
Clone a slide framework your coding agent can author. open-slide is an MIT React runtime with a fixed 1920 × 1080 canvas, presenter mode, comments embedded in source, and export to static HTML, PDF, and editable PowerPoint. It uses pnpm, and its agent workflow depends on the included slide skills.open-slide
Awesome LLM Apps is a curated collection of 100+ AI Agents, Agent skills, and RAG apps. It covers models from OpenAI, Anthropic, Google, and open-source models like GLM, DeepSeek, and Qwen that you can run locally on your computer. (Now accepting GitHub sponsorships)
That's all for today. Come back tomorrow for the next batch of AI tools, model drops, agent repos, and weird benchmarks worth your time.
If you found one thing to try, share the issue with someone who ships.
Some teams never seem to stop moving. They're on Attio, the agentic CRM.
Every customer signal is captured in one shared context layer, always current and compounding. Agents and workflows build pipeline, chase every buying signal, and move deals forward, an always-on revenue engine running alongside your team.
With Attio, you’ll get:
- Leads automatically prioritised and routed to the right rep
- Expansion and risk signals caught the moment they land
- Follow-ups written in your voice, already there when you arrive
Teams like Parallel, Turbopuffer, and Wordsmith build on Attio. Are you one of them?