AI Tech
This section gathers materials on tooling for AI agents: memory, skills, MCP servers, saving context. They were written by me with AI from the projects' documentation and repositories. They will help you get to know these tools and pick the ones you need for your own task.
Agents
AgenticSeek: An Agent Entirely on Your Own Machine
The only agent here that never calls a provider: six specialists with a dispatcher run on a model on your own machine. An honest table of graphics-card requirements comes with it - and it is sobering.
Agents
Crush: A Terminal Agent That Reads Code Through a Language Server
An agent from the people behind half the pretty console interfaces takes its context from the same place your editor does - a language server. Configuration is ordinary Bash with its own builtins, and that has a non-obvious security consequence.
Agents
Aider: A Repo Map, Automatic Commits, and a Year Without a Release
The project that invented the repository map and automatic commits before anyone else: instead of files in the context, a concise scheme of the whole codebase with signatures. The ideas spread across the industry; the tool itself has gone a year without a release.
Agents
Cline: One Engine in the Editor, the Terminal and the Board
Five products on one engine: a library, a command line, a VS Code extension, a JetBrains plugin and a web board. Project rules are written once and work everywhere - but tool auto-approval in the terminal is on by default, and the front page promises the opposite.
Token savings
claude-code-router: One Entry Point for Every Agent
A local gateway between ten agents and their providers: the agent talks to one address while the model, account and fallbacks are decided in one place. Subagents pick the cheap model themselves from descriptions, and every request shows what it cost.
-768x512.webp&w=3840&q=75)
Token savings
ccusage: The Agent Bill, Counted from Local Logs
This topic explains how to cut spending but never showed how to count it. This command reads the logs of sixteen agent CLIs straight from disk and turns them into reports by day, session and five-hour window - with no install and no data leaving your machine.
-768x512.webp&w=3840&q=75)
Extensions
MCP Servers: Seven Reference Implementations, Not a Catalog
The protocol's official repository holds not a server catalog but seven reference implementations and a warning that these are educational examples, not finished solutions. Thirteen others were archived with no security guarantees, and two of them still appear in the configuration example.
-768x512.webp&w=3840&q=75)
Harnesses
SuperClaude: Not Software, but Instructions in the Context
The project calls itself behavioral configuration rather than software: thirty commands and twenty agents exist as Markdown files in your home directory. Structure and a command vocabulary come in exchange for context window, and the documentation names the figure itself.
-768x512.webp&w=3840&q=75)
Harnesses
wshobson/agents: One Set of Skills for Six Harnesses
One Markdown skill set works in six harnesses: adapters build native artifacts for each. The unit of installation is the plugin rather than the skill, and the catalog comes with a rare thing - a table honestly listing what each harness loses.
-768x512.webp&w=3840&q=75)
Process
Task Master: Tasks as a File the Agent Reads Itself
A product description turns into a task list with dependencies and statuses that the agent queries itself: asked what comes next, it answers with the first open task rather than a guess. Inside: a licence with a condition, per-branch tags, and recounted tool sets.