Releases·Tool Release
Open-sourcing our agent toolbox: FastMCP server fleet + Matryoshka semantic memory (25k vectors in 220ms on CPU) + Git worktree isolation + llms.txt onboarding
A team open-sourced its internal agent toolbox, including a FastMCP server fleet, a Matryoshka semantic memory system (25k vectors queried in 220ms on CPU), Git worktree isolation, and llms.txt-based onboarding. The release gives practitioners reusable components for building agents.
1 engagement1 sourceTue, Sep 8, 2026, 03:01 PM
AI Overview
- The team open-sourced an agent toolbox built around a fleet of FastMCP servers
- Matryoshka-based semantic memory retrieves 25k vectors in 220ms on CPU
- Agent sessions run in isolated Git worktrees to avoid conflicts
- An llms.txt file onboards agents to the codebase and tools
Related
Social·Community
Liad Yosef on MCP Apps: agents ignore llms.txt and have no brand loyalty
A talk by Liad Yosef, co-creator of the MCP Apps spec, reports that while roughly half of tested websites publish llms.txt files for agents, almost none of the agents tested actually read them. Yosef argues the agentic web should be designed from observing thousands of agent runs rather than written specifications, and that MCP Apps — which lets servers return real interfaces into chats — unlocked the shift. He illustrates the stakes with his own team switching analytics vendors after their coding agent preferred a competitor's better API.
A talk by Liad Yosef, co-creator of the MCP Apps spec, reports that while roughly half of tested websites publish llms.txt files for agents, almost none of the agents tested actually read them. Yosef argues the agentic web should be designed from observing thousands of agent runs rather than written specifications, and that MCP Apps — which lets servers return real interfaces into chats — unlocked the shift. He illustrates the stakes with his own team switching analytics vendors after their coding agent preferred a competitor's better API.
44 engagement1 sourceyoutubeWed, Sep 16, 2026, 02:30 PM
Social·Community
MiniCPM5-2B: The Best Sub-Agent Model Yet?
A YouTube review by Sam Witteveen examines openbmb's MiniCPM5-2B, a 2B-parameter open model positioned for sub-agent use. The video covers its benchmarks, a 'JustRL' training recipe, and a demo, alongside a Runpod serverless deployment walkthrough.
A YouTube review by Sam Witteveen examines openbmb's MiniCPM5-2B, a 2B-parameter open model positioned for sub-agent use. The video covers its benchmarks, a 'JustRL' training recipe, and a demo, alongside a Runpod serverless deployment walkthrough.
424 engagement1 sourceyoutubeThu, Sep 10, 2026, 01:00 PM
Social·Community
H Company releases Holo4, an open-weight 27B agent model that clicks GUIs, runs code, and calls MCP tools
A YouTube video by Sam Witteveen reviews Holo4, H Company's open-weight model family that operates a GUI, writes and runs code, and calls MCP tools within one agent loop. The video links H Company's blog and Hugging Face weights, showcasing the release to developers.
A YouTube video by Sam Witteveen reviews Holo4, H Company's open-weight model family that operates a GUI, writes and runs code, and calls MCP tools within one agent loop. The video links H Company's blog and Hugging Face weights, showcasing the release to developers.
384 engagement1 sourceyoutubeTue, Oct 6, 2026, 02:10 PM