Opening
Kiri has only been live for a month, and it already hit 8.4k stars on GitHub. The official benchmark claims container startup takes just 2.5 milliseconds, nearly 80x faster than Docker Desktop. That's a pretty brutal number. In today's open-source roundup, besides breaking down how Kiri pulls this off, we'll also look at how AgentRun helps AI agents burn two-thirds less tokens, plus day-trader-skill, a day-trading workflow built from 11 chained Claude Skills, and I'll show you how it keeps the agent from going rogue and placing orders on its own.
Today's Open Source Picks
1. Kiri (libracoder/kiri): A lightweight container engine aimed squarely at Docker Desktop's pain points
- What it is: Docker Desktop is expensive, heavy, and eats up a fair amount of resources. Kiri is a lightweight container engine written in Rust, built to replace both Docker Desktop and OrbStack while staying fully compatible with your existing docker CLI, Compose, and buildx workflows. It's for engineers who run containers locally on a regular basis.
- Recent traction: 8.4k stars on GitHub (a new project within the past 30 days, observed on November 16, 2026 during backfill production, not broadcast-day historical traction).
- Key features:
- Containers started via the Docker Engine API take around 2.5 milliseconds, nearly 80x faster than Docker Desktop (213ms) and considerably faster than OrbStack (380ms) as well
- Local domains
<name>.kiri.localcome with built-in HTTPS via a self-managed CA, essentially baking in a free ngrok alternative - VMs auto-sleep after 60 seconds idle, the daemon's resident memory footprint is only about 15MB, and wake-up takes roughly 0.15 seconds
- One command
kiri k8s startspins up a single-node Kubernetes cluster, andkiri importlets you migrate images and volumes from OrbStack or Docker Desktop with a single command
- Quick start: Download the installer for macOS, Windows, or Linux (.dmg/.msi/.deb/.rpm) from runkiri.com. Once installed, switch your docker context to kiri and keep using your usual
docker runanddocker compose upcommands as-is, no habits to relearn. - Compared to peers: Versus Docker Desktop, Kiri is fully open source and free (Apache-2.0), with no extra licensing fees for enterprise use. Versus OrbStack, the official benchmark puts Kiri's API response at 2.55ms, well ahead of OrbStack's 380ms.
- Who it's for: Engineers doing container-based development every day who can't stand how slow and resource-hungry Docker Desktop is.
- Link: github.com/libracoder/kiri
2. AgentRun (parcha-ai/agentrun): Stop making your AI agent relearn everything from scratch every time
- What it is: Does your AI agent re-explore the same APIs and re-read the same docs on every single task, burning through tokens in the process? AgentRun is a platform that a YC team open-sourced after using it internally to cut their own costs. It layers on top of existing agent harnesses like Claude Code and Codex, adding self-learning, resumable execution, and cost controls.
- Recent traction: 138 points on Hacker News, with a 40-comment discussion thread.
- Key features:
- Cross-run self-learning: after each run, the agent proposes notes, skills, or workflows worth keeping, which get committed to git once they pass code checks and an AI grader review, then get reused automatically next time. The official case study shows per-run code review cost dropping from 0.0077
- Durable execution: all runs and steps are stored in the same database, so if the server restarts mid-run, execution picks up where it left off. A 30-second lease mechanism lets other runners take over tasks that disconnect
- Built-in spend caps that auto-pause execution, plus per-tool allow/ask/deny settings, so the agent can pause and ask you a question, with the run waiting for your answer
- Quick start: Run
npx @parcha/agentrunlocally with a single command, and it spins up the server plus runner (data stored in ~/.agentrun) and automatically opens a browser control panel. Log in with ChatGPT or supply a model key to try the built-in templates. - Compared to peers: Versus running Claude Code or Codex CLI bare, AgentRun adds cross-run learning, resumable execution, and a cost dashboard, positioning it somewhere between a raw agent harness and an enterprise-grade agent platform. And since the core is fully open source under Apache-2.0, you can self-host it rather than relying on a closed SaaS.
- Who it's for: Engineers already using Claude Code or Codex for automation who are tired of burning tokens re-learning the same things every run.
- Link: github.com/parcha-ai/agentrun
3. day-trader-skill (jakeschincariol/day-trader-skill): A complete day-trading workflow built from 11 chained Claude Skills
- What it is: This isn't an automated order-placing trading bot. Instead, it breaks down the day-trading workflow into 11 clearly divided Claude Skills, covering everything from scanning the market and tracking news to calculating entry/exit points, simulating win rates, and sending order alerts. It's entirely free, MIT-licensed, requires no API key, and serves as a great example of a multi-skill collaboration architecture built on Claude Skills.
- Recent traction: 8.1k stars on GitHub (a new project within the past 30 days, observed on November 16, 2026 during backfill production, not broadcast-day historical traction).
- Key features:
- 11 skills working in a relay: scanner finds opportunities → news monitors headlines → entry/exit sets entry and exit points → timemachine simulates win rates → budget/risk manages position sizing → broker places orders → alerts sends notifications. Each skill reads and writes to the same
trader/folder to hand off work - Time Machine resamples the stock's 5-minute candles from the past 60 trading days, runs the same strategy 10,000 times in simulation, and calculates the probability of hitting take-profit, stop-loss, or staying open, giving you an honest win rate before placing an order instead of relying on gut feel
- Dual order-safety mechanisms: paper trading is enabled by default, only the trader-broker skill can place orders, it requires you to explicitly type "yes" in the conversation, and there's an additional Allow once/Deny confirmation layer built into Claude itself
- 11 skills working in a relay: scanner finds opportunities → news monitors headlines → entry/exit sets entry and exit points → timemachine simulates win rates → budget/risk manages position sizing → broker places orders → alerts sends notifications. Each skill reads and writes to the same
- Quick start: Paste the repo link into Claude and say "install skill", or run
/plugin marketplace add Jakeschincariol/day-trader-skillfollowed by/plugin install day-trader-skill@day-trader-skillin Claude Code. - Compared to peers: Unlike typical trading scripts that connect directly to a broker API for automated order placement, this one deliberately keeps the agent from trading autonomously, emphasizing paper trading and dual human approval. Architecturally, it's also a much more complete demonstration of the multi-skill collaboration and file-handoff design pattern in Claude Skills than a single-prompt trading bot.
- Who it's for: Engineers interested in studying multi-agent collaboration architectures built on Claude Skills, or anyone who just wants to backtest a day-trading strategy using public data.
- Link: github.com/Jakeschincariol/day-trader-skill
Closing
All three tools today are really pointing at the same thing: the open-source community is no longer satisfied with single-purpose tools. It's now open-sourcing entire capabilities, performance, self-learning, multi-skill collaboration, that used to require paying for. If you can only try one today, I'd start with Kiri. It has the lowest switching cost, and you'll feel the speed difference immediately. Come back on November 17, 2026, and I'll dig up the next batch of open-source finds for you.


























Comments