BACK_TO_FEEDAICRIER_2
LocalCowork puts MCP agents on-device
OPEN_SOURCE ↗
REDDIT · REDDIT// 36d agoOPENSOURCE RELEASE

LocalCowork puts MCP agents on-device

Liquid AI has introduced LocalCowork, an open-source desktop agent powered by LFM2-24B-A2B that runs fully on-device and uses MCP tools for file operations, OCR, security scans, and audit logging. The bigger story is that Liquid is positioning local, privacy-first agents as practical on consumer hardware, with reported tool-selection latency around 390 ms on an M4 Max.

// ANALYSIS

LocalCowork matters less as a polished consumer app than as a concrete proof that local agent UX can feel fast enough to be useful. Liquid’s argument is straightforward: for privacy-sensitive workflows, sub-second tool routing on a laptop beats shipping data to the cloud for a smarter but slower agent.

  • The open-source repo ships a large tool surface: 75 tools across 14 MCP servers, with a curated 20-tool demo set tuned for better accuracy.
  • The strongest claim is responsiveness, not autonomy: Liquid reports 80% single-step tool-selection accuracy, but only 26% end-to-end completion on multi-step chains.
  • That tradeoff makes the product more compelling as a human-in-the-loop workstation agent than as a fully autonomous operator.
  • The local audit trail, no-cloud design, and support for security/document workflows give it a real angle for regulated or privacy-sensitive teams.
// TAGS
localcoworkagentmcpopen-sourcedevtool

DISCOVERED

36d ago

2026-03-06

PUBLISHED

36d ago

2026-03-06

RELEVANCE

8/ 10

AUTHOR

Zc5Gwu