Open Source Radar — August 24: agents that learn you

Today's board has a clear thumbprint: the top of trending is dominated not by another chat wrapper but by agent harnesses that build a memory of you, ways to run coding agents for free, and the prompt libraries and plugin markets feeding them. Five fresh projects, none of them repeats from the week's earlier radars.

Hermes Agent (Python, ~235k stars) — Nous Research's open-source agent is the top mover on today's board, and its pitch is "the agent that grows with you." Unlike most harnesses that start cold every session, it runs a built-in learning loop: it creates skills from real experience, sharpens them as you use them, nudges itself to persist what it learns, and searches its own past conversations rather than making you re-explain the same context. It runs on whatever model or provider you already pay for instead of locking you into a single one. If you want an assistant that accumulates competence — not just a longer chat history — this is the lead pick on the board.

free-claude-code (Python, ~48k stars) — A local proxy that lets you run Claude Code, Codex, and other coding agents against a catalog of about 50 provider models, including free tiers worth over a billion tokens a month, without putting your accounts at risk. It's built to stay friendly to each provider's terms of service and drops any integration the moment a supplier stops allowing it, and it works from a terminal, an editor, or your phone, with local voice input. The angle is that the expensive, subscription-gated agents are often the pleasantest to use — so this pulls their front-ends down to free or cheap models. It's the biggest star mover besides the frameworks, and for anyone tired of paying for every coding-agent subscription, hard to argue with.

OpenHuman (Rust, ~37k stars) — A personal AI "brain" that is local-first and aimed squarely at getting to know you in minutes rather than weeks. On a first sync it compresses your inbox, calendar, documents, and messages into a memory graph so the agent starts with full context instead of a blank slate, and it routes tool output through a compression layer to keep that ambition affordable. It's also an orchestrator rather than a single loop: agents talk to each other over end-to-end encrypted messaging, and long-running work is checkpointed so a stuck task gets steered and a halted one reports a root cause. It's early and GNU-licensed, but it topped GitHub trending for nine straight days on launch — a sign the "agent that knows you" bar is finally what people are chasing.

awesome-gpt-image-2 (JavaScript, ~12k stars) — A prompt library for OpenAI's GPT-Image2 that treats prompts as code: more than 470 reverse-engineered cases and a couple dozen reusable templates, split into atomic pieces — subject, lighting, materials, layout — so you can build stable, controllable, repeatable image styles instead of sketching prompts by hand. It's built to be consumed by agents and automation, and it ships the same style catalog as a reusable skill your coding agent can pull in, so the exact library the website uses is what lands in your workflows. You'd reach for it when a GPT-Image2 generation needs to look intentional and consistent, not lucky.

Claude Plugins Community (Python, ~1k stars) — Anthropic's read-only mirror of the marketplace plugins for Claude Cowork and Claude Code. Every entry has been submitted by a developer, passed automated security scanning, and been approved before it lands, so you're getting a curated directory rather than a wild-west list. This one is less about novelty and more about the signal it sends: as multi-agent adoption grows, plugin marketplaces — with vetting built in — are becoming the distribution layer humans trust to extend their agents.

Worth watching this week: the framework race is shifting from "which model" to "which memory" — the repos winning stars right now are the ones that learn who you are.

If your agent could keep learning about you across every app, would you let it — or is starting fresh each time actually cleaner? Tell us in the comments.

Sources: Hermes Agent (GitHub) · free-claude-code (GitHub) · OpenHuman (GitHub) · awesome-gpt-image-2 (GitHub) · Claude Plugins Community (GitHub)