The Pull: Week 16, 2026
This week in open source: Prototype restraint, long-term AI memory, and a flexible client that keeps you out of vendor handcuffs
1) Andrej Karpathy Skills: Stop Letting AI Overengineer Your Prototypes
github.com/forrestchang/andrej-karpathy-skills
Featured in The Push: April 13th, 2026
Andrej Karpathy Skills is a specialized configuration file that forces Claude to stop overengineering code and start acting like a senior developer. It uses clever reasoning constraints to ensure the AI asks questions before making assumptions and only touches the code it is supposed to. This is for anyone tired of AI-generated bloat.
6290 stars/day | 8.8 watchers/day
2) Hermes Agent: The Assistant That Actually Remembers You
github.com/NousResearch/hermes-agent | License: MIT
Featured in The Push: April 13th, 2026
Hermes Agent is a self-improving assistant that uses a closed learning loop to turn past experiences into permanent skills. It is clever because it decouples the AI from your local machine, allowing for persistent, cross-platform conversations via Telegram or Discord. Anyone who is tired of re-explaining context to AI should look.
4296 stars/day | 2.8 watchers/day
3) Thunderbolt: The Anti Lock-in AI Client
github.com/thunderbird/thunderbolt | License: MPL-2.0
Featured in The Push: April 18th, 2026
Thunderbolt turns AI chat into deployable infrastructure instead of a vendor-owned tab. The clever bit is the local-first sync architecture plus provider abstraction, which makes model choice and data control feel operational, not theoretical. Teams dealing with compliance, procurement, or multi-model workflows should pay attention.
781 stars/day | 5.0 watchers/day
4) Evolver: Prompt Tweaks Need Governance
github.com/EvoMap/evolver | License: GPL-3.0
Featured in The Push: April 17th, 2026
Evolver turns agent improvement into a governed protocol instead of a pile of prompt edits. What stands out is the asset-based architecture, where reusable behavior changes and audit logs become part of the system itself. Teams running long-lived AI assistants, especially in ops-heavy or compliance-sensitive contexts, should pay attention.
742 stars/day | 3.5 watchers/day
5) OpenAI Agents Python: Agent Frameworks Get Real
github.com/openai/openai-agents-python | License: MIT
Featured in The Push: April 18th, 2026
OpenAI Agents Python turns agent building into a structured runtime, with delegation, tracing, sandboxed workspaces, and approval flows in one framework. The smart part is treating agents like operational systems, not fancy prompts. Worth a look for anyone building voice assistants, internal copilots, or workflow-heavy AI products.
904 stars/day | 2.0 watchers/day
6) Claude Code Game Studios: AI Needs Middle Management
github.com/Donchitos/Claude-Code-Game-Studios | License: MIT
Featured in The Push: April 17th, 2026
Claude Code Game Studios turns one Claude Code session into a structured game team with roles, approvals, and review paths. The smart part is the governance model, not the agent count, because it makes AI output behave more like coordinated studio work. Worth a look for indie teams, game founders, and anyone tracking AI-native workflow software.
830 stars/day | 2.5 watchers/day
7) Voicebox: ElevenLabs, But Keep Your Data
github.com/jamiepine/voicebox | License: MIT
Featured in The Push: April 14th, 2026
Voicebox turns local voice cloning and speech generation into an actual desktop production workflow, not just a one-shot AI demo. The clever part is the queue-and-version architecture that makes multi-engine generation, long scripts, effects, and editing feel stable on your own hardware. Creators, indie studios, and teams building privacy-sensitive voice features should look.
785 stars/day | 1.0 watchers/day
8) Claude Mem: AI Memory Should Be Default
github.com/thedotmack/claude-mem | License: Other
Featured in The Push: April 14th, 2026
Claude Mem turns Claude Code into a system with memory across sessions, using compression, hybrid retrieval, and context reinjection instead of brute-force transcript stuffing. The clever part is the layered recall model. Anyone betting on AI assistants for repeated project work should look closely.
1559 stars/day | 2.8 watchers/day
9) Generic Agent: Agents Should Grow Skills
github.com/lsdefine/GenericAgent | License: MIT
Featured in The Push: April 16th, 2026
Generic Agent turns one-off agent work into a growing library of reusable skills. What makes it interesting is the layered memory architecture, which stores compact operational knowledge instead of hauling giant transcripts back into context. Worth a look for anyone tracking AI agents, browser automation, or software that gets better through repeated use.
632 stars/day | 1.0 watchers/day
10) Dive Into Llms: The AI Course People Actually Finish
github.com/Lordog/dive-into-llms
Featured in The Push: April 15th, 2026
Dive Into Llms turns LLM education into a modular, hands-on map of the actual AI stack, from tuning and reasoning to jailbreaks, watermarking, multimodal systems, and agent safety. The clever part is the curriculum design itself. Founders, PMs, students, and investors trying to build real AI judgment should look.
802 stars/day | 3.5 watchers/day
11) Magika: File Detection Finally Grew Up
github.com/google/magika | License: Apache-2.0
Featured in The Push: April 16th, 2026
Magika turns file type detection into a fast, calibrated ML system instead of a shaky extension check. The smart part is its confidence architecture, which knows when to return a precise label and when to back off. Security teams, storage products, and any company handling messy uploads should look.
523 stars/day | 1.3 watchers/day
12) Pascal Editor: Figma For Buildings Is Here
github.com/pascalorg/editor | License: MIT
Featured in The Push: April 14th, 2026
Pascal Editor turns browser-based building design into a real product foundation, not just a flashy 3D demo. The smart part is the architecture: flat scene nodes, selective geometry updates, and reusable editing systems. Founders, product teams, and anyone building spatial software should pay attention.
430 stars/day | 0.6 watchers/day
13) Open Agents: AI Coding Needs a Back End
github.com/vercel-labs/open-agents | License: MIT
Featured in The Push: April 16th, 2026
Open Agents turns AI coding into a durable cloud workflow instead of a fragile browser chat. The smart part is separating orchestration from execution, so runs can persist while sandboxes hibernate, resume, and safely touch real repos. Founders, platform teams, and anyone building AI devtools should pay attention.
251 stars/day | 0.3 watchers/day
14) Deep Gemm: The AI Speed Layer
github.com/deepseek-ai/DeepGEMM | License: MIT
Featured in The Push: April 18th, 2026
Deep Gemm turns low-precision GPU math into a coherent kernel layer for modern LLM workloads, not just a one-off speed demo. The clever part is the runtime JIT plus unified handling of scaling layouts, grouped workloads, and fused MoE execution. Worth a look for AI infra teams, model labs, and anyone obsessed with GPU efficiency as product strategy.
215 stars/day | 1.0 watchers/day
15) Kronos: Reading Stock Charts Like Open Source Poetry
github.com/shiyu-coder/Kronos | License: MIT
Featured in The Push: April 13th, 2026
Kronos turns messy financial charts into a structured language that a Transformer can read. It uses a clever tokenizer to predict future price movements by treating them like the next word in a sentence. It is a must-look for anyone building the next generation of automated trading strategies.
450 stars/day | 1.3 watchers/day
16) Apollo 11: Software History, Not Nostalgia
github.com/chrislgarry/Apollo-11 | License: Other
Featured in The Push: April 15th, 2026
Apollo 11 turns the moon landing’s guidance software into a public, reviewable repository instead of a frozen legend. The clever part is the verification model: GitHub is used not just to host historic code, but to preserve provenance and improve fidelity over time. Anyone interested in systems design, software history, or trustworthy technical archives should look.
170 stars/day | -0.8 watchers/day
17) AI Hedge Fund: Finance Twitter, Compiled
github.com/virattt/ai-hedge-fund
Featured in The Push: April 15th, 2026
AI Hedge Fund turns stock analysis into a panel of AI investors, then runs the result through risk and portfolio logic instead of stopping at a chatbot answer. The clever part is the workflow design: disagreement is structured, signals are stored, and final decisions get constrained. Fintech builders, market research teams, and curious investors should look.
370 stars/day | 1.2 watchers/day
18) Craft Agents: AI Needs a Better Workspace
github.com/lukilabs/craft-agents-oss | License: Apache-2.0
Featured in The Push: April 17th, 2026
Craft Agents turns AI agents into a persistent workspace with shared sessions, reusable skills, connected sources, and remote execution. The clever part is making agent capability live at the workspace level, not inside one chat. Teams exploring AI ops, internal tooling, or founder workflows should pay attention.
73 stars/day | 0.0 watchers/day