Blog
-
Designing tools that AI can actually use
The tool description is a big part of the interface. If the model can't tell when to call a tool from its description and schema, quality falls apart.
-
Loading code without the disk: what each OS lets you get away with
memfd, PE loaders, unlink-after-dlopen. Three operating systems, three different definitions of 'in memory'.
-
paker: load encrypted Python packages from memory
Encrypt a Python package tree, ship it over the network, load it from memory. What works, what doesn't.
-
I built a read-only MCP server for Akamai
16 tools for searching CDN properties, browsing EdgeWorker code, querying DNS, and inspecting network lists. All read-only.
-
Your coding agent is a slot machine
AI coding agents compressed the feedback loop from hours to seconds. That compression has a name.
-
Read-only MCP servers: what they are and why I only build them
A read-only MCP server can read but never write: no create, update, or delete. Why that removes an entire class of agent failures, and how I actually enforce it.
-
I benchmarked MCP servers in 5 languages
Python, Rust, Go, TypeScript, C#. When does the language actually matter?
-
Your MCP server is not an API adapter
MCP servers can cache, pre-aggregate, and transform data independently of upstream services. Here's how and why.
-
The parts of your agent that aren't the model
The model provides the reasoning. Everything around it determines whether it works once or every day.