Today brought a wave of releases covering both tooling and real-world operational data around local LLMs. This round-up focuses on llama.cpp's full MCP support, measured power efficiency numbers for the AMD MI50, a hands-on report on building an agent environment with 4GB VRAM, and the SEC's procurement of AI surveillance agents.

llama.cpp Now Supports All MCP Protocols

The MCP integration led by ngxson is now complete, with [PR #26062](https://github.com/ggml-org/llama.cpp/pull/26062) merged. HTTP-based MCP servers could already be handled on the client side, but running stdio servers required real integration work — and that gap has finally been closed.

By switching `llama-cli` to a server-based path, MCP support has been layered on top of the existing native tool server. As a result, the llama.cpp WebUI can now be used as a fully-fledged agent chat. MCP servers can be configured via a standard JSON config file or specified inline on the command line.

The example given is a setup that connects a coding-oriented MCP server such as [Serena](https://github.com/oraios/serena). It appears you can now build an agentic coder that runs on local models alone with no external dependencies — a long-awaited milestone for anyone who has been waiting for a fully local agent environment.

AMD MI50 Power Consumption Tests: 50W Is the Sweet Spot

A report posted in the LocalLLaMA community on Reddit measured inference performance on an AMD MI50 (gfx906) while capping power consumption. The numbers come from a practical configuration: Qwen3.6-35B-A3B at IQ4 quantization with a 256K context.

The most practical takeaway: at a 50W cap, the card delivers 70% of peak performance on just 26% of the power draw. Throughput per watt comes out to 0.458 t/s/W, 3.6× the 0.173 at 190W. Tightening down to 20W raises efficiency to 6.0×, but prompt processing speed falls to 53% of peak, so the report's practical recommendation lands at 50W.

Another finding: prompt processing is more sensitive to power limits than generation speed. Generation depends on memory bandwidth while prompt processing is compute-bound, so throttling hits them differently. Even at 100W the card delivers 97.5% of peak performance, so anyone building a dedicated inference box may find it worth revisiting their TDP settings.

A Hands-On Report on Building a Local Agent with 4GB VRAM

Also on LocalLLaMA, an interesting hands-on report covers building an agent environment on an RTX 3050 Ti Laptop (4GB VRAM). The setup runs on an i7-12700H / 32GB RAM / Ubuntu 24.04 and is built around local Qwen models (the author self-hosts Bike4Mind, a BUSL-1.1-licensed workspace product).

Speeds per model are as follows:

qwen3.5:0.8b — 122 tok/s, ~1.4GB, fully on GPU

qwen3.5:2b Q4_K_M — 96 tok/s, ~2.4GB, fully on GPU (the sweet spot)

qwen3.5:4b — 25 tok/s, ~3.4GB, one-third spills over to CPU

qwen3.5:9b — 8.6 tok/s, almost entirely on CPU

The 2b fits within 4GB and outperforms the previous qwen2.5-coder:3b (72 tok/s), making it the current optimal choice.

That said, using it as an agent hits plenty of walls. With RAG in the picture, the embedder (qwen3-embedding:0.6b, ~1.2GB) and the chat model fight over VRAM, forcing a model swap every time. Artifact generation with the general-purpose qwen3.5 produces issues like unclosed tags, requiring a switch to qwen2.5-coder. The report also notes that enabling many tools leads to routing mistakes — limitations that highlight the constraints unique to small models.

Report: The SEC Is Moving to Strengthen Surveillance with AI Agents

Making the rounds on Hacker News was a report that the U.S. Securities and Exchange Commission (SEC) is procuring AI agents to monitor phone location data, social media, and credit headers. The source is a Substack article by Jack Poulson. Discussion on HN is just getting off the ground, so see the original article for details — but from the standpoint of expanding civilian surveillance, this is a topic worth keeping an eye on.

Wrap-Up

On the local front, llama.cpp's full MCP support looks set to significantly reshape future agent setups, while the 4GB VRAM case study neatly captures the current reality: the tools are coming together, but you still have to live with the quirks of small models. The MI50 power data should serve as a useful reference for weighing TDP settings for dedicated inference machines against real numbers.