CoreCoder

You will get more from reading this than running it, and that is a compliment. It is small enough to hold in your head in an afternoon, which is the deciding trait, because every argument you have about agent design gets sharper once you have seen one that fits on a long train journey. It genuinely works, which is what separates it from a tutorial.

Pick it if you want that understanding, or a base to build your own thing on. Pick Claude Code when the job is shipping today and you do not need to know how.

A blacklist is a list of the dangerous things somebody remembered. It stops the obvious commands and it cannot stop the ones expressed differently, which is a well-understood limitation the project states plainly rather than hiding. The problem is the pairing: shipping a mode that executes a prompt without a human present, guarded only by pattern matching, puts the weakest control in front of the least supervised path.

What it does right: it asks for consent before anything mutates the disk, so in normal use the human is the gate rather than the regular expression.

1. Context management is the substantive contribution: compaction in three tiers rather than one truncation strategy, which is the mechanism most agents implement badly and document not at all. 2. Cost is reported per run in tokens and currency, making efficiency measurable instead of anecdotal, which almost nothing else on this board does. 3. Omitting tool protocols and retrieval is a scoping decision the author states outright.

No benchmark is claimed and none is needed. The observation: an agent that reports what it spent is more useful to research than one that reports what it scored.

There is no business here and there was never meant to be one. The eight accompanying essays are the actual product, and their currency is reputation rather than revenue, which is a perfectly good trade for an individual and a terrible one for anybody planning around it. Seventeen hundred stars in five months says the explanation landed.

Likely path: the author is hired by somebody who read it, and the repository becomes a portfolio entry. Position: no company to hold a view on, and the code is worth more to your team as reading than as a dependency.

One sentence on the demo: it worked and it was small. Then the questions that end these conversations. There is nobody to call, nobody to sign anything, and no commitment that next month's version will exist, which means standardising sixty engineers on this makes us the maintainer of a tool we did not write. No identity integration, no central policy, no audit record.

The onboarding cost is genuinely low, and I would rather spend it having two engineers read the source than deploying it. Not yet, and not as a standard.

This is the shape I keep asking for and rarely get. Permissive licence, one package command, and pointing it at the model server on my own machine is a single environment variable rather than a plugin, with an optional router if I want to reach hosted providers through the same interface. Subagents are there when a task splits.

Best of all it was written to be forked, and the author says so, which removes the usual argument about whether modifying somebody's project is rude. Nothing here holds state I cannot reach and nothing phones anywhere I did not point it.