_flux

jq cannot be just an MCP, unless it's acceptable that yuo pass all data through the context. If that's not acceptable and you want to have it as tool, then you need some other way to handle the data.

I think the basic solution to this is to have a "static shell" but with modern tools for the agents, not actually executing other binaries. It could have things like jq, curl, piping and redirection to/from session files. Maybe even Python if it can be made safe. If not, then there are a lot of languages can be.

As far as I know, the only way an MCP can provide you data that doesn't go into the context is by providing URLs to the data, and then the model uses e.g. curl to access that data for data manipulation purposes. You could also return result set ids and provide accessors to such data, but ultimately you'd need to provide powerful accessors to that result set to avoid polluting context. Things like shell with all its power already provides.

It seems like there's little point in MCP in that case. Maybe more point if it was a standard mechanism for MCP to provide additional data, in a completely compatible fashion with all other tools? You could perhaps even pass such URLs to other MCPs. You could have an MCP for jq for doing stream processing. Starts to look a lot like a shell, though.

Seems like MCPs could also be extended to store auxiliary data to your memory (or filesystem..), and then an additional extension to provide that kind of data as auxiliary data in the calls to MCP.

Well, even as is, MCP still provides a standard method of using OAuth for accessing such services. And you must use MCP if you wish to add something to the ChatGPT.com web service, so it's easy to see why OpenAI folks are seeing companies going that way.

Hmm, indeed, so maybe I could have all this as an MCP, so I can just easily pass any imaginable data manipulation inside it, and then also have it support calling other MCPs, all inside that one MCP, to avoid filling context with intermediate data..

Sounds a lot like a shell to me.

Not having access to the shell is a big hindrance. I have my agent access Gitlab and Jira via CLI tools and in so many cases jq or python is used to manipulate or combine the data into a more useful format, making use of pipes and temporary files. You can of course limit what an agent can do, most easily by not giving it access to things it shouldn't do. I suppose there are no existing easy gateway methods to grant fine-grained OS-level permissions to add such things back, except perhaps `sudo` and similar tools.

MCPs are impossible to combine this way: everything you feed or get from them goes though the model and consumes tokens.

Virtual terminal is surprisingly capable, it even has word delete feature! And shells have some nice history features which I'm sure would become even more familiar if one had to use them (as I suppose was the case a long time ago), such as !$ that expands to the last argument of previous entry.

But they are just convenience features, not in my opinion the defining property. They basically work just the same without them.

On the other hand, there are many elements in the user interface that are not text. In fact, I think it would be quite understandable if one even separated REPLs like bash or octave to be "text user interfaces" while applications that make use of character placement and border special characters "terminal user interfaces", because they use means beyond text stream to communicate with the user. One might even render straight up graphics to a terminal [emulator application].

Even X had a separate application called xterm 42 years ago: the complete X system was not to my knowledge called a terminal system, except perhaps when discussing the dedicated client devices, such as VT1300. Also the term "virtual terminal" as far as I know has always referred to a the kind of interface this application is making use of.

So I think we can just accept that the term is overloaded such that "terminal" refers to both of these situations, as there is no historical precedent to have it exclude the other situation, and the term "terminal-based application" is completely clear to a rational listener.

It would be so nice if these "well typed languages" would be able to interact better with each other. Alas, the GC will probably cause some friction, e.g. the ocaml-rs package makes this pretty visible: https://github.com/zshipko/ocaml-rs/blob/master/examples/rus... . (I have not used either of these projects.)

Maybe the GC would be less of a problem if we didn't want to have the best possible performance as well.

Many years ago Maemo (the mobile OS from Nokia) had the profile manager (day mode/night mode etc) written in Prolog. To me it seems like it's a very appropriate application for it.

To me it would certainly make sense if the protocol just said "append this text to context window id/sha256", in particular as the data is cached in tensor level in the provider side, so they need to first do that lookup anyway. So I would be surprised if they don't have that.

In addition, this protocol could make it more transparent to say "oh we cannot proceed as we dropped the this cache, are you sure you want to proceed and consume a whole lot of expensive uncached tokens?". Oh, maybe that's a reason not to do it..

So, this website being called "Hacker news", means.. ?

While hacking is used to to refer to illicit activities, I do actually believe that the same activities can also be performed e.g. on your own devices, or with permission, and still be called hacking. So in my view, I do not believe legality is the defining term, but the actual things you do; oftentimes useful for illegal activities.

But I must assume there is a group of people who consider the term to be loaded with that exclusion, so I should take this into account.

However, to me "hacking your own devices" as a concept seems crystal clear, and is not a contradictory term.

I actually don't think that's possible, if you assume the objects in the video can also shake and the intent is to eliminate only the camera shaking.

I think gyroscopic data still can have the edge if it has higher sampling rate than the video: then it could be used for removing blur from individual frames.

I also expect purely gyroscopic approach to be much lighter compute-wise.

It clearly seems people have different meanings to the word, then.

For example, if I am able to gain root access to a WiFi access point I own, even though the vendor has tried to prevent it, then yes, I would call it a hack. To me, it doesn't matter why or who is doing the steps.

In fact, I believe I have never before heard someone combine the meaning of the word to be related to the ownership of the device being hacked.

I suspect the number of people understanding the word in your way is a minority. Redefining terms doesn't help build mutual understanding: here we are taking a word some think has negative connotations and then remove the thing they think should be cool and ok, and then suggest that this is actually the real meaning of the word. Personally I don't think this is how words should be wielded.

I have no trouble believing it is better :), but it is not as easy to mirror a Github issues, or CI configuration, to Forgejo or back as it is to handle the git side.

I think Radicle is interesting. It doesn't solve the CI bit, at least not yet, but I suppose it's possible to hook up some local runner for it.

There's also a bug tracker which I believe was called bug, but I can't find it ;), that tries to bridge different issue trackers and providing offline mode for working with them.

People of course also love free CI capacity where they can run even untrusted code, so in that sense Microsoft resources might be difficult to compete against.

I think you're forgetting issue tracking and CI.

I imagine some (many?) people just enjoy accessing their printer over the mobile phone without setting up VPN, even if said mobile phone isn't on the home wifi.

Often I find LLMs doing multiple steps to achieve some goals (e.g. do certain operations against JIRA or Gitlab), and if the LLM work seems useful, I instruct it to create a tool to achieve the task more directly and revise skill data to make use of the tool.

Granted I've let it mostly vibecode those tools, so they might be garbage. I should perhaps have it do a refactoring round to make more composable tools..

Just because something is good at finding bugs, it may not find all the bugs. Finding a bug only tells you there was one bug you found, it doesn't tell if the rest is solid.

In addition LLMs also make bugs, and debugging assembler is more difficult, wasting more tokens, thus more money.

A very big practical reason is also that assembler code would eat context like no other.

> GPT for instance had a lot of issues using git worktrees, and didn't understand how to correctly use it to then merge stuff back into a main branch, vs Claude which seems to do this much more naturally.

I wonder how much of that is due to the model being somehow better, or the harness having built-in instructions on how to use them.

I've used worktrees with Codex just fine, but I instructed it to use my scripts for setting it up and tearing it down. The scripts also reflinked existing compilation artifacts to speed up compiling and allocated a fresh db instance for it, but then also applied a simple protocol for locking the master repository during merges, so multiple agents wouldn't try to merge at the same time. It has been following those instructions quite well.

I've played a bit with having a Claude generate a TLA+ model, I review it, Claude reviews it, and then Claude checks for alignment with the actual system. The code in the system can be annotated with comments to link back to the model, and vice versa, to make human and AI review easier.

I think it's highly promising. I might even call it a success. For example, I was able to express a bug in the system and Claude was able to find me a property that would reveal that in the model; and then revise the model to eliminate that bug, and then revise the code accordinly.

Additionally (with OpenAI) I have a small guest management system I had the LLM generate a TLA+ model about the API which endpoints a client can request from it, and which tokens it learns from it (there are session ids, event ids, guest ids and invitation ids that interact in interesting ways, as guests are not required to have an invitation to the system, and they can have only magic-link protected account, resulting in different visibility on the contact details of others), and what data it can or cannot access with those tokens.

I still haven't manually verified, but it seems rather promising as well.

I think it could potentially be useful. Sometimes I need "simple" shapes that still are somewhat annoying to create. And I think you don't need to one-shot these, the process is permitted to be iterative! The skills can be improved by time by revising AGENTS.md, e.g. "when I say L-bracket, I probably mean..".

I think going from a picture to an initial starting point with well-"thought"-out structure for CAD purposes could potentially be very useful. Optimally you could just enter the measurements and be done.

..and what kind of tracking, e.g. anonymous usage statistics vs update checks, e.g.

  *:analytics=1:google_analytics=0,syncthing:upgrade=1

The specification could go on and on!

I always choose to go with positive terms with variables etc, so this would then be ALLOW_TRACKING=0. It brings in some consistence and makes it easier to reason, as you get to avoid double negation.

Perhaps the "DO NOT TRACK" name is somewhat of an established term, though.

I think it's possible to come up with a situation where you want to do a sorted search per every pixel in the screen, for every frame.

I think the difference is that Cloudflare is the one party providing streaming access for their customers: not just anyone can proxy the data through Cloudflare, they need to be a Cloudflare customer first.

When I'm posting this message to Hacker News, I'm the "customer" of this website. I'm not customer of all the intermediate nodes in the chain. So if I were to write something illegal and HN would be irresponsive to takedown requests, the courts could order the IP of HN to be blocked, not some intermediate ISP.

I'm relatively sure the source code I've written and stored in my local computer is not sucked up in the LLM training data. And I believe people working with Blender models are pretty much in the same situation: they don't host their data in a third-party service and openly share it.

How could that possibly work?

Let's say we didn't have assemblers, but instead we would have three professions:

- Specifiers, who make the specification for the system

- Programmers, who write C code

- Machine encoders, that take that C code and write machine code for a CPU

Would it be that the copyright would then belong to programmers, if no other explicit assignments would be made?

---

Thinking about it, probably yes: copyright of the spec belongs to specifies, copyright of the C belong to programmers, and copyright of machine code to machine encoders. Or would it depend on the amount of optimizations the machine encoders would do, i.e. is it creative or not? And then does this relate to the task and copyrightability of C compiler output, where optimizations can sometimes surprise the developer?