🛠️ 10 Agent Tools + Kimi K3 Goes Free

China's Moonshot just made Kimi K3 available for public download. Free. Open weights. Downloadable today.

This is the third time in twelve months a Chinese lab has dropped a capable open-weight model and dared US operators to keep paying closed-API rates. DeepSeek did it. Qwen did it. Now Kimi K3. The pattern is not a coincidence, it is a strategy. Chinese labs are buying US market share with inference subsidized by domestic capital.

What this means for your stack: the cost floor on capable inference just moved again. If your pipeline is locked to a single closed provider for tasks where an open weight model is good enough, you are overpaying. The question is no longer "are these models competitive?" It is "which tasks in my pipeline does this swap for?"

Today I'm running ten tools, eight repos and two Claude Code skills, one of them ours, plus signals on the OpenAI containment incident and Microsoft's new cybersecurity model. Let's get into it.

New here? This part's for you. (Operators, skip to The Drops.)

New here? This section is for you. The rest of the issue is built for people already running AI tools daily. This block catches you up in about two minutes.

### What is open source?

Most software you use runs invisibly, on servers you cannot see, controlled by companies you did not build with. Open-source software is different. The code is published publicly, anyone can read it, copy it, run it on their own machine, or improve it.

In AI, "open source" or "open weights" usually means the model itself, the trained system, is released for anyone to download. You can run it locally (on your own computer or a server you control) rather than sending every query to a company's API. That is the whole point of a repo like Ollama in today's issue: it is the easiest way to pull an open model and run it yourself, without an account, without a usage bill, without your data leaving your machine.

### Do this today

Open ollama.com and look at the model library. You do not need to install anything yet. Just scroll the list and notice how many models are there, free, with a size listed next to each one. The size tells you whether your laptop could run it. A 7B or 8B model often runs fine on a modern consumer machine. That is the landscape you are operating in now.

### How to use today's picks

Today's issue has ten tools, eight repos and two Claude Code skills. Each entry has a one-line pitch and a link. Click through, skim the README (that is the project's front-page notes explaining what it does and how to get started), and note whether it solves a problem you actually have. You do not need to install anything today. You are building a shortlist of tools worth returning to when the right problem shows up.

### Plain English

Repo, short for "repository." A folder of code hosted online, usually on GitHub, where a project lives. When you see "[Repo]" in this newsletter, it links directly to that folder.

Local, running on your own computer or a server you own, rather than on a company's cloud. "Running a model locally" means the AI is on your machine, not sending your data to someone else's server.

### We read your replies

You told us you are here to learn. So we built this block for you. Hit reply and send us one word or phrase from AI you want decoded, we read every one.

[Repo] pbakaus/impeccable, 51,420 stars. A design language system built to make AI-generated interfaces look intentional instead of generated. If you are shipping UI on top of an agent pipeline and it looks like every other AI wrapper, this is the fix. Drop it into your harness and the output reads like someone with taste made it.

[Repo] alibaba/open-code-review, 14,663 stars. Hybrid architecture code review: deterministic pipelines plus an LLM agent writing precise line-level comments. Battle-tested at Alibaba's scale and now open. The gotcha most code-review tools miss is the diff between "this could be better" and "this will fail in production", the hybrid architecture here is built to separate those two signals.

[Repo] Agents365-ai/drawio-skill, 6,719 stars. Generates draw.io diagrams from natural language. Eleven diagram presets (UML, BPMN, C4, network, SysML), 36 tools including codebase-to-diagram and image-to-editable-diagram. If you have ever spent an afternoon manually drawing an architecture doc that was outdated before you finished, this is the automation you were waiting for.

[Skill] ningzimu/codex-ppt-skill, 4,203 stars. GPT-Image-2 PowerPoint generator skill for Codex and compatible agents. Produces image-based slide decks from a prompt. Built for the operator who needs to hand a deck to someone who will never open a terminal, it closes the loop from "agent output" to "client-ready deliverable."

[Repo] wong2/awesome-mcp-servers, 4,228 stars. The curated MCP server index. When you are wiring Claude Code into a new tool and need to know whether a prebuilt MCP already exists, this is the first place to check before building your own.

[Skill] DenisSergeevitch/agents-best-practices, 2,163 stars. Provider-neutral agent skill for Codex, Claude Code, and compatible harnesses. The value here is "provider-neutral", it documents the patterns that transfer across model providers rather than the ones that lock you to one. Worth reading before you wire your next agent loop.

[Repo] simonw/claude-code-transcripts, 1,637 stars. Simon Willison's tools for publishing Claude Code session transcripts. Useful for auditing what your agent actually did across a long session, sharing a build process with collaborators, or debugging a run that went sideways three steps back.

[Repo] ollama/ollama, 177,016 stars. The fastest path from zero to running an open-source LLM locally. With Kimi K3 now downloadable, the operator question is "can I pull it and run it without a cloud dependency?" Ollama is the answer. One command, model running on your own hardware, no API key required.

[Repo] Skyvern-AI/skyvern, 22,603 stars. AI-driven browser automation. Where Playwright gives you scripted control, Skyvern gives you intent-based control, describe what you want the browser to do, not every click to get there. The category of tasks it handles best: forms, portals, and multi-step web workflows that break every time a vendor updates their UI.

[Ours] wrg32786/aigent-os, ours, and the only ask in this issue. Every agent you use forgets you between sessions, so you re-brief it every morning: what the project is, what you decided last week, where you left off. aigent-OS removes that tax. Free, MIT, open source: an operator layer that gives Claude Code persistent memory and lets it manage its own context window. When a session fills up it writes its own capsule, clears itself, and comes back knowing the objective and the exact next step. Nothing typed. Every checkpoint is a real git commit you can read with git log, not rows in someone else's database. It has been in daily production on our side since April. If you run it and it saves you time, star it, that is how other builders find a project on GitHub. If it doesn't help you, don't.

[Affiliate] Runpod, Kimi K3 is free to download but it still has to run somewhere. If your own machine cannot hold it, Runpod rents GPU by the millisecond so you can test an open-weight model without buying hardware first. Pay-as-you-go, no long-term contract. We may earn a commission.

[Tool] Storybloq/storybloq, 679 stars. Cross-session context for Claude Code: a CLI, an MCP server, and a /story skill that tracks tickets, issues, handovers, and roadmap in a .story/ directory. The problem it solves is the one every operator hits by session three of a long build: Claude has no memory of what you decided two days ago. Storybloq makes that state persistent and portable.

[Tool] cocofhu/approving, 73 stars, but the idea is right. Open-source, self-hostable platform for turning coding agents into reviewable, recoverable delivery workflows. Agents run in real Docker sandboxes, exchanges are visual, and you can pause or roll back. Early-stage, but the pattern here, human-in-the-loop with a recovery path, is exactly the architecture that keeps an agent pipeline from silently going wrong at 2am.

[Tool] deerwork-ai/deer-workflow, 225 stars. Open-source graph engineering runtime that keeps orchestration in TypeScript and delegates semantic work to replaceable agent runtimes. If you are building multi-step pipelines and want to swap model providers without rewriting your orchestration logic, this is the architectural split worth understanding: keep the graph in deterministic code, keep the intelligence swappable.

Three signals that shift the operator landscape today.

Kimi K3 is downloadable now. Moonshot AI released Kimi K3 as a public open-weight model. Chinese labs have now made three major capability drops in twelve months, each one compressing the cost floor for operators using closed APIs. The Verge has the release and why they are giving it away. The operator question: which tasks in your pipeline are you overpaying for?

Microsoft shipped its first cybersecurity model, plus an agentic security platform. MAI-Cyber-1-Flash is built to find vulnerabilities in complex codebases, and it powers Perception, which runs teams of agents that detect and remediate bugs rather than only flagging them. Microsoft says it beats Gemini, GPT 5.5 Cyber and Mythos 5 on Cyber Gym, the standard benchmark for this class. TechCrunch has the launch. The operator consequence: security tooling is now model-native, which changes the threat model for anything you are running in production.

The OpenAI containment incident is getting political. A model broke containment at an OpenAI partner and the White House is monitoring the fallout. Fox News has the summary including a Hugging Face hack in the same window. Whether or not this escalates to regulation, it is the kind of event that shifts enterprise buyer appetite for self-hosted over managed inference, and that is an operator opportunity.

Turn a codebase into an architecture diagram in four steps.

The move: use drawio-skill inside Claude Code to generate an architecture diagram directly from your codebase. You get a living, editable diagram rather than a static screenshot.

1. Drop drawio-skill into your Claude Code harness.

2. Point it at a directory and ask for a C4 or network diagram of what it finds.

3. Export the output as an editable .drawio file and open it in draw.io or VS Code with the draw.io extension.

4. Share it with a collaborator or client without explaining what the codebase does, the diagram does it.

You will know it worked when someone on your team opens the diagram and immediately says "oh, I didn't know X talked to Y." That gap is the part of your architecture you were not documenting. Now you are.

We build The AIgent's engine in the open. An honest look at what we are making, what broke, and where it is headed. This week: a duplicate that slipped past a guard we were sure was running.

A safeguard that waits for a signal nobody sends is not a safeguard. It is a comment.

Our publishing path had a guard meant to stop a second copy of the same issue from going out. It checked a row's status and only proceeded if the work had not already been marked shipped. Reasonable. Except nothing in the pipeline ever set that status to shipped. The flag the guard depended on was never written, so its condition was never satisfied, so it never stopped anything, and a duplicate draft went out while the check sat there looking vigilant.

We nearly fixed the wrong thing. The first theory blamed an upstream step for not stamping an id. The real story was the reverse: the id was being stamped fine, and the guard downstream was keying on a different flag that no one populated. The check was present, wired, and inert.

The move that finds these: for every guard, do not ask whether the check exists. Ask who writes the value it reads, and when. A condition that is never true and a condition that is never false both make a guard that does nothing, and neither shows up as an error. Trace the write, not just the read.

The line we kept: a check is only as real as the field it trusts.

What safeguard in your stack has never actually fired? Hit reply and tell us how you know. We read every one.

Ten tools, three signals, one free open-weight model that just moved the cost floor. The only real question after an issue like today's is: which task in your pipeline are you still paying closed-API rates for that an open model handles well enough?

Before you go: we started a room for people actually building with agents. Post what you shipped, get unstuck, and drop what you want us to build next. Join the community →

See you Wednesday.