Back to blogs

About the author

xbill

@xbill

Master of MCP and herding cats

New York

From the author

Articles by xbill

Explore all insights

Explore xbill's latest articles and ideas.

Sep 21

Find the AWS Resources Nobody Is Using, and What They Cost You

A local-first CLI that scans an AWS account for unused resources, prices each one from the AWS Price List API, and generates or performs the cleanup. Read-only by default, with every price fetched from the live API and dated in each report.

9 min readRead article

Sep 21

Put the Arithmetic in the Tool: an MCP Server for an AWS Waste Scanner

A cost report gets read twice: once by a person in a terminal, once by an agent through an MCP server. Making the server compute every total instead of handing back rows keeps the two answers the same, and measures the report's own arithmetic against the AWS bill.

8 min readRead article

Sep 20

One Iceberg MCP Server, Seven Catalogs: What It Takes to Reach Each One

Step by step: one Python MCP server with four read-only Apache Iceberg tools, pointed at Polaris, BigLake, OneLake, Glue, S3 Tables and Horizon by changing one environment variable. All four tools work on all six. What changes per catalog is the login, the storage package, and one Azure credential that takes 553 seconds.

7 min readRead article

Sep 20

One Iceberg MCP Server, Seven Catalogs: What It Takes to Reach Each One

Step by step: one Python MCP server with four read-only Apache Iceberg tools, pointed at Polaris, BigLake, OneLake, Glue, S3 Tables and Horizon by changing one environment variable. All four tools work on all six. What changes per catalog is the login, the storage package, and one Azure credential that takes 553 seconds.

7 min readRead article

Sep 18

Gemma 4 on a Tesla T4: QAT Weights Decode 1.79x Faster Than bf16

A step by step deployment of Gemma 4 E2B with vLLM on a single Tesla T4 attached to a Compute Engine VM, and a measured comparison of the QAT w4a16 checkpoint against the bf16 reference on the same card.

9 min readRead article

Sep 18

It's Got What Content Craves: Sanity, Built for the People of Idiocracy

An agent proposes watering the crops, a human Cabinet approves in a picture-button Studio, and Sanity Workflows waters and harvests. Plus a stopwatch on three ways to edit the same data.

9 min readRead article

Sep 17

Serving Gemma 4 on an AMD MI300X: What $1.99 an Hour Buys

A step by step deployment of Gemma 4 E2B to a single AMD Instinct MI300X on AMD Developer Cloud, driven by Python MCP tools, and the throughput a 191.7 GiB card returns for its hourly rate.

9 min readRead article

Sep 17

An MI300X Over MCP: What the Matrix Cores Execute, and What They Don't

One AMD Instinct MI300X on AMD Developer Cloud, managed entirely through a tag-scoped Python MCP server, with every figure read off the card rather than a spec sheet. fp8 e4m3fnuz runs 1.77x bf16; int8, which AMD rates identically to fp8, runs 0.69x; fp4 is not on this silicon at all. One droplet, $1.99 an hour, and two readings that were wrong the first time.

20 min readRead article

Sep 16

A 4 GB Laptop GPU Beats a 12-Core CPU by 4.3x on Gemma 4

Serving Gemma 4 E2B q4_0 through llama.cpp on one laptop, twice: CPU-only and on a 2021-era 4 GB GTX 1650 Ti. Same GGUF, same binary, same prompts, one flag apart. The card takes decode by 4.3x, and needs only 1598 MiB to do it.

6 min readRead article

Sep 16

Four Iceberg Tools, Three Agent Frameworks: What Ports, and What Doesn't

Four read-only Apache Iceberg tools bound into Google ADK, AWS Strands and Microsoft Agent Framework, run against five catalogs, 360 timed runs. Building the agent ports and running it does not; speed follows the model and how much it writes; the storage wiring under the tools is the per-cloud work.

18 min readRead article

Sep 16

Four Iceberg Tools, Three Agent Frameworks: What Ports, and What Doesn't

Four read-only Apache Iceberg tools bound into Google ADK, AWS Strands and Microsoft Agent Framework, run against five catalogs, 360 timed runs. Building the agent ports and running it does not; speed follows the model and how much it writes; the storage wiring under the tools is the per-cloud work.

18 min readRead article

Sep 15

Nano Banana 2 Lite in Kiro CLI 3: MCP 2.0, the New Interactions API, and Headless Permissions

The Kiro edition of the Nano Banana 2 Lite MCP server, updated: FastMCP is now MCPServer, google-genai 1.x gets a 400 from the Interactions API, and Kiro CLI 3 needs a permissions rule before it will call the tools headless.

12 min readRead article