Waku Agent open-source local agent framework passes 2,000 stars

Shen Sean Chen @ShenSeanChenOriginally posted 2026.10.14

Open original on X

𝗪𝗮𝗸𝘂 𝗔𝗴𝗲𝗻𝘁 (@waku_agent) 𝗷𝘂𝘀𝘁 𝗰𝗿𝗼𝘀𝘀𝗲𝗱 𝟮,𝟬𝟬𝟬 𝘀𝘁𝗮𝗿𝘀 𝗼𝗻 𝗚𝗶𝘁𝗛𝘂𝗯 𝗶𝗻 𝟯 𝗺𝗼𝗻𝘁𝗵𝘀 (𝗥𝗲𝗽𝗼: https://git.waku.one). Here's what 3 months of building in public taught me: 1. Agents people trust are legible. If you can read a loop's tool usage, memory retrievals, and evals, you can fix it. A customer told me what matters most to them is that he can check an agent's work in 30 seconds. 2. Memory is the part people poke at hardest. One customer told me initially that it saved too much, too often. And I hit the other problem myself: it lived on one laptop and never reached my other agents. 3. Contributors show up when the code is small enough to hold in your head. They sent bug reports and proposals for Windows, a Milvus store and a Slack gateway, and one rebuilt our architecture whiteboard as code. 4. Memory gets harder in a team. My agents read my team's commits, and my team's agents read mine. What we really share is context. Waku Agent is a local-first AI agent harness your team actually own. Loop, memory and evals all in code that stays readable as it grows. 2,000+ stars, 398 forks, and 31 people who have shipped code into it. Thank you to everyone who starred, forked, filed an issue or sent a PR. That's where http://waku.one, our memory product, comes from. First, your memory shouldn't reset every time you try a new agent. New agent drops, Grok Bot, Muse, Hermes or whatever's next, you connect http://waku.one over MCP and it starts with what the last one learned, not a blank page. Then the team: shared spaces where a decision one teammate's agent learns gets reviewed and reaches every teammate's agent. When a newer decision replaces an old one, Waku retires the old fact instead of handing it back. Waku Agent stays open source, and it's still the easiest way to see the memory at work. Try Waku Memory: https://waku.one If it's useful, a star helps more than you'd think.