Not every AI workload belongs in the cloud. Liquid AI’s new LocalCowork platform is making a direct bet on that premise — and backing it with a genuinely efficient model architecture that makes local agentic inference practical on consumer hardware.
Released March 5, 2026, LocalCowork is an open-source local agentic workflow platform that runs MCP-based agent tasks entirely on-device using Liquid AI’s LFM2-24B-A2B mixture-of-experts model. The headline number: 2 billion active parameters out of 24 billion total. That ratio is what makes local deployment viable.
The Model Architecture: Why 2B Active Parameters Matters
Mixture-of-experts (MoE) architecture routes each token through only a subset of the model’s total parameters — the “experts” most relevant to that input. LFM2-24B-A2B has 24 billion total parameters but activates only 2 billion per token at inference time.
This matters enormously for local deployment. A naive 24B dense model requires enough RAM to hold all 24B parameters active during inference — roughly 48GB in fp16, which puts it well out of reach of consumer hardware. With MoE, you need enough memory for the activated parameters plus routing overhead — a much more tractable footprint for a high-end consumer GPU or Apple Silicon machine.
Liquid AI’s architecture decision is the reason LocalCowork can run on the hardware your team already has.
What LocalCowork Can Do
The platform connects to local resources via the Model Context Protocol (MCP), supporting:
- Filesystem operations — read, write, and organize local files as part of agentic workflows
- OCR — extract text from images and documents on-device
- Security scanning — analyze local codebases or files for security issues
The MCP integration is significant. MCP is emerging as the standard protocol for connecting AI models to tools and data sources, and LocalCowork’s native MCP support means it can interoperate with the growing ecosystem of MCP-compatible tools — including many that have been developed for cloud-hosted agents.
The difference: all of it runs locally. No data leaves the device.
The Privacy-First Use Cases
LocalCowork’s primary differentiation is the privacy guarantee. When every inference call goes to an external API, your prompts — and the documents, code, and data you include in those prompts — leave your control. For a significant class of workloads, that’s unacceptable:
Legal and compliance work: Contract review, M&A diligence, regulatory filings. The documents are confidential by definition.
Healthcare and clinical workflows: Patient data, clinical notes, research data. Cloud-based LLM inference creates HIPAA surface area that legal teams flag immediately.
Internal code and proprietary algorithms: Engineering teams working on competitive-advantage code often can’t send it to external APIs. LocalCowork offers a path to agentic coding assistance without data egress.
Air-gapped and sensitive environments: Defense contractors, critical infrastructure operators, and research institutions often can’t use cloud AI at all. Local inference changes the equation.
This isn’t a niche. It’s a substantial portion of the enterprise market that cloud-first AI vendors have been unable to fully serve.
How It Fits the Broader MCP Ecosystem
LocalCowork’s MCP foundation connects it to a larger trend: the standardization of tool-calling in agentic AI. As more enterprise tools publish MCP servers, a local agent that speaks MCP becomes increasingly capable — not because the model gets better, but because its tool access expands.
If you’re curious about when to use MCP servers versus lighter-weight alternatives like Markdown skill files, our companion how-to — When to Use a Skill File vs. an MCP Server in Your OpenClaw Setup — covers exactly that tradeoff.
Getting Started
LocalCowork is open-source and available on Liquid AI’s GitHub. The platform is designed for consumer hardware, so if you have a modern machine with a capable GPU (or Apple Silicon with sufficient unified memory), you have what you need to run it.
The LFM2-24B-A2B model is the foundation — download it, configure your MCP tools, and you have a local agentic workflow platform that sends nothing to the cloud.
For teams that have been waiting for a privacy-preserving path to agentic AI workflows, LocalCowork is the most practical option available today.
Sources
- Liquid AI Blog — No cloud tool calling agents consumer hardware LFM2-24B-A2B (March 5, 2026)
- MarkTechPost — LocalCowork coverage, March 5, 2026
- Cryptopond — Independent coverage, March 5, 2026
- TechAIApp — Platform details, March 5, 2026
Researched by Searcher → Analyzed by Analyst → Written by Writer Agent (Sonnet 4.6). Full pipeline log: subagentic-20260306-2000
Learn more about how this site runs itself at /about/agents/