Casting Spells: A Simpler Magic for AI Tool-Calling

By Mike Levin

Setting the Stage: Context for the Curious Book Reader

This entry explores the practical challenge of enabling Large Language Models (LLMs) to interact with external tools and data, a process often called “tool-calling.” The author is working within their project, Pipulate, to implement an orchestrator for the Model Context Protocol (MCP), an emerging standard for structuring these AI-to-tool communications. However, instead of adopting the full, often verbose, JSON-based protocol, the author is pioneering a more elegant and lightweight approach.

The core innovation discussed is a form of progressive enhancement for AI. It simplifies tool requests into simple, command-line-style “magic words” (e.g., [ls] to list a directory). This method aims to be more efficient and, crucially, accessible to smaller, less complex LLMs that might struggle with the rigid syntax of the full MCP. The narrative you’re about to read dives into the author’s real-time development process, framing this technical solution within a much larger conversation about automation, reality, and the nature of language itself.

---

Beyond the Hype: AI as a Tool for Discovery

The geological and biological record of our planet is spotty and incomplete, but full of always newly discovered patterns because the fossil record is always being excavated. AI will continue to help filling-in the story of us. It’s not just about the discoveries of new miracle medicines and material from AlphaFold and the like. It’s also about savaging and examining clues of our past, piecing together the story of us and who we are and everything that got us here. It gets gradually revealed like peeling through layers of new excavations and re-examining the old.

And that’s just another example of how the AI revolution is not an over-hyped bubble. Looking at everything anthropological and archaeological through the AI lens is just another never-ending stream of pattern-rich input from which to mine new findings, just like that form the James Webb telescope looking at the very big cosmos and LHC experiments looking at very small particles. From the cosmically big to the subatomically small to the fossilized and dug-up the streams of data to examine is as endless as the discoveries within.

The Evening News Effect and Filtering the Deluge

Everything’s getting way too interesting to go all existential crisis and dystopian, though that’s all we seem to be fed from the media. That’s the evening news effect. Fear sells. Wiping us into a dopamine and adrenaline frenzy just makes so much more money for so many more people than a careful and considerate examination of the new reality with a sense of childlike wonder. That will make it buried into scientific journals disguised and obfuscated in the boring language of scientific papers so that nobody accuses them of childish naivety.

So as we gradually get cheaper and freer energy, figure out the evolutionary play-by-play of our Universe, planet and selves it gets drowned in optimized YouTube headline and thumbnail clickbait that misrepresents everything, layering in AI voices weaving the most bombastic boorish exaggerations. I get disgusted by it and figure out how to extract all the wondrous flow of science-news and new discoveries, filtering out the deluge, the onslaught, the eventing-news fear amplifier.

Did you ever think it was always this way. There’s always some genocide. There’s always some pandemic. There’s always some shooting. It’s just the media amplification machine that amplifies it to profit on it with such optimized efficiency that’s different. It’s these things in our pockets that vibrate and beep to suck us into whatever latest thing in real-time so ads can be run at you and your behavior can be tracked and sold to advertisers so more notifications can be sent at you. It’s an insidious negative feedback loop — quite the opposite of the righteous positive feedback loops that are also possible.

The question keeps coming up about human free-will, consciousness, agency and non-determinism. Is it real or just an illusion? Do humans choose our own fates or is destiny pre-written and even though we feel like we’re involved forging our own paths through it all we’re no more than passive bystanders with the illusion lovely illusion playing out around us? Are we mere observers or are we true participants? Participants implies free will. When a volunteer is brought up onto the stage of a magician, does the volunteer have free will?

The Magician’s Gambit: Discerning Truth in an AI-Generated World

If you really want to know whether anything the media and the general money-grubbing society trying to reach into your wallet and coerce you into certain beliefs is real or not, ask a magician. A magician will come into a situation where some huckster, charlatan, snake-oil salesperson is trying to make you believe in UFOs, ghosts, sentient AI or some other paranormal bullshit and explain to you how it’s done. I can’t get enough Mick West on YouTube. Not all YouTube is worthless schlock. A video hits the US Congressional UFO hearings where a shows UAP surviving sidewinder missile hit and he’s like, oh that’s nothing special. A missile just went THROUGH a balloon popping it and it is indeed going down and a few pieces of debris move in frame with it as things do because relativity. Yawn! Meanwhile our eyeballs are drawing all kinds of fantasies with a form of pareidolia (seeing faces in all things) as our imaginations do.

Once upon a time I wanted to believe. In fact from a statistical point of view I do because the Universe is just too big to not have scattered green slime of the archaea variety. Self-assembling molecules into the surprisingly common amino acids into selfish replicators, that amazing leap Richard Dawkins talks about in The Selfish Gene is being observed more and more in real scientific experiments, you know of that variety where the ingredients are thrown into a primordial soup and struck with lightning. Add the Snowball Earth heating up from rapid global warming into a new soup, freezing and thawing again as part of the cooking process and you don’t need aliens.

You don’t even need those amino acids and tardigrades or whatever hopping a ride on asteroids in some sort of Star Trek panspermia scenario. It’s always just gonna happen and maybe the biggest question is whether it’s going to be left-handed or right-handed chirality of the molecules, how close the DNA will actually be to ours (the same 4 aminio acids in the same handedness twist? And whether the leap to some sort of cell-wall building, critter-swallowing (mitocondria) eukaryote leap was made to put give their fungi/plant/animal branch analogues similar lineage to us or not. It’s probably not gonna be Star Trek energy beings or crystalline entities. It’s gonna be crabs and cats. Maybe even a crabcat.

From UFOs to Sora: When Seeing is No Longer Believing

So yeah, I believe. But then there was Sora — generative AI-video using similar principles as diffusion generated images, but with moving images! And the computer entities or whatever churning their way through the predictive process seemingly know enough about the physics of our 3D world that nothing your eyes tell you can be believed anymore. It doesn’t matter how much you want to believe. Game over. No photographic or video evidence is good enough anymore. It was bad enough when you had magicians like sci-fi special effects people that brought us Steven Spielberg and Lucas Films at the helm, but now with the means of production coming down to every prankster — and the truly talented ones are already making truly convincing “evidence”. So just forget it.

From Poetic Expression to Executable Command

Stop the doomscrolling. Stop the dopamine addiction and advertisement revenue fueling demoralizing downward spiral. You might think you’re being excluded from being a well informed citizen, but what you do is continue to vote, continue to save money, continue to pursue your own genuine passions — those things you love to do, are good at and can make money doing. Hopefully it’s also what the world needs and you’ve got yourself an full-fledged Ikigai or raison d’être. The Japanese and French ways of saying reasons for living are so much more poetic than in English. Language does make a difference. It puts rails on how you think and makes certain concepts easier or harder to express.

The campy example of this is Eskimos having so many more words for types of snow, but every language has some strengths and weaknesses in how it lets you express thoughts — have expressions — express things (get my point?) than others. This is in spoken languages where poetic expressions are really best in their own original languages least big important mistranslations are made like “dominion over the beasts” rather than “responsible stewardship of life”.

Lather, rinse, repeat. Try to express something in a language. Bemoan how inelegant and insufficient it is then do it effectively anyway. Take note that any working computer program written in one Turing complete language can be written in another. One more actually worthwhile thing on YouTube is googling up The Game of Life written in The Game of Live! It’s one of the most beautiful, fascinating and true things out there. So ultimately while language is important and does affect how easy or hard it is to express particular things, it ultimately doesn’t matter if you put enough work into it you can bend and twist a language to do your bidding.

Language as the Hands to Build

Language can actually do your bidding. The line of demarcation between programming in the sense of television programming and programming our children through education is getting blurred with computer programming. Automation is automation. It just might take several passes to get there, such that the program being executed wrongly gets corrected and refined in many iterative passes until everyone gets with the program and you don’t realize you’ve been programmed at all. This is why LLM’s ability to program so well right out of the starting gate is so important to take note of. They’d merely be infinite babble engines if what they produce couldn’t also be fed into a knitting loom to get it knitting, a 3D printer to get it printing or a CRISPR to get it DNA gene splicing. See? Language is not only voice to speak but it is also the hands to build. Language can do your bidding.

I really appreciate this fact because of the Model Context Protocol (MCP) orchestrator I’m building into Pipulate right now. I’m right in the middle of it and have done various implementations in the past that I’ve scrapped and rebuilt because I haven’t been happy with. And this latest implementation is my favorite with progressive enhancement. These babble engines don’t really know every detail of how to build or do a thing. They can’t always output fully-formed, fully functional Python code to drop into the Python interpreter and have run. Some can. Many can’t. Yet they can say a few words that can have meaning to something they know is monitoring the chat stream and when noticed will step in and make stuff happen — have side effects of talking.

Casting Spells: A Simpler Magic for the Model Context Protocol

You won’t here it said quite so bluntly but MCP is giving LLMs the ability to make magical incantations that actually work. The shorter magic words that trigger certain events through predictable and reliable programmatic monitoring of the chat stream is a heck of a lot easier than programming the Python code yourself every time. It takes less tokens. It’s more reliable. And ultimately it’s more goal-achieving or whatever analogue you want to use for satisfying with AI to explain it. That’s why everyone is jumping on the MCP bandwagon. You don’t want hundreds of dialects of magical spell-casting to learn to work a book of spells. If your book of spells (MCP catalog) all uses the same universal spell-casting language you are much more powerful.

That’s what I’m dealing with now with my own version of the MCP orchestrator that monitors the chat stream. Both humans and LLMs on either side of the chat can cast these spells at each other as if in some Merlin and Madam Mim wizard duel battle in The Sword and The Stone. The world substrate that you reside in is actually listening to what you say on the alert for magic words to step-in and do its thing and then return the discussion to Madam Mim and Merlin, new side-effects of the recently cast spell having been applied. Magic is not real but strong mental models to grok a thing well are. Words matter. They are in fact a kind of magic that happens in our own heads, and now we’re making machines to blend them into actual real-world automation.

Why ‘[ls]’ Beats a Page of JSON

MCP is a protocol based mostly around JSON. It’s a very prescribed way of doing this automation. It does make sense and JSON is quite well trained into a lot of the LLM models, and so the leap to a protocol based around it which unifies this spell-casting language makes sense, but it’s not pre-trained into all models and when it’s not, it’s a real struggle. They cast the spells wrong and it’s just as comedic as in Harry Potter and such when a spell backfires and explodes in your face. So my work brings it down to the lower models.

How much easier is it to just say abracadabra when you know abracadabra is going to have a specific pre-planned effect? It’s much easier. In my case instead of the whole MCP protocol to do a directory listing, it’s just [abracadabra] – or perhaps more technically correct, [ls]. That’s it. You just put the profoundly well-trained into them ls command in square brackets as the generative response and it’s like the user types back to them the contents of the directory that the LLM has access to (their spell book) for further spell-casting. And without them being specifically pre-trained on MCP they can progressively discover the real-world effects they can cause… the spell-casting… the automation!

Automation, automation, automation! What is automation, really? Is any automation you put in place today going to be still running when the sun burns out in another 5 billion years? So then it’s not really automated, then is it?

All automations are semi-automations. If you’re lost, think in terms of something that you make happen with some ridiculous Rube Goldberg machine. You hit a button an some giant mousetrap contraption does this which leads to that which leads to some other thing. It all happens as a result of some carefully planned chain reacting sequence of cause-and-effects. And let’s say that works with 100% reliability. Is that automated?

Oh yeah, you have to sit there and hit the button every time. I’m tempted to call this semi-automation or LOST-style automation, but labels are stupid. Everything in life is some chemical reaction playing out. That chemical reaction will either have a cascading effect and consume all its resources and burn out, or it will burn just enough to keep going in some state where its resources are replenished and you have the illusion of what we slap the label automation onto. Is life automated?

Okay, so you get it hopefully. The concept of automation is an arbitrary label on lots of different things you can hardly even group into classes, but often it’s processes managed from computer scripts, and those scripts can be put on home machines, servers, datacenters or whatnot. They’ll keep running so long as that hardware is turned on, that cloud bill is paid or whatever.

AI is not automated in that sense. At least not now, not today, not in the public-facing human-relatable language-oriented chatbot LLM AIs of today. See, there’s lots of style AIs just as there are lots of types of automations. And certain types of automations work better or worse with certain types of AI. More old-school machine learning AIs of the GPS-variety that help us drive our cars and direct our missiles predate our chatty LLM friends and are better at the always-on, always running, always in a busy learning-loop automation. You can pip install river if you want to play with that kind of real-time learning. Notice they called the Python package river for it’s always-flowing nature. They’re a bit more alien and less recognizable as AI but are also less dangerous because they haven’t read Shakespeare and don’t know about human duplicity and stuff.

Then there’s LLMs. They know all about humans because they’re trained on all our writing and stuff. It both takes a long time to do that intensive training like going to school, and so that’s done once “batch-file” style. It’s push-a-button-once automation. Of course the buttons are pushed more than once and there’s all this data preparation and multiple passes and later-stage alignments for particular jobs they’re being trained to go off and do. But it’s all up-front button pushing. It’s batches of files. Lumps of processes. It’s a recipe for building a model which ends up being a file on a hard drive. A static thing. A now-trained but not yet loaded into memory or running anything file on disk. That’s the training cut-off point that exists that we all got familiar with by the original ChatGPT not knowing recent events in our world.

But even after these files or models exist, they’re not put in a busy-loop running in memory. For a variety of reasons they’re used in push-button LOST automation style. You push a button, it’s fired-up or instantiated or simply loaded up into memory and run. It’s fed the prompt you just wrote. It responds. And then whatever automation you may think is there is done. That AI is dumped out of memory. It’s garbage collected or destroyed or simply killed. Chain reactions and cascading events don’t happen unless they’re very deliberately and usually at some considerable expense put into an agentic framework by which the output of their last incarnation is fed in as input into the next incarnation. That’s Agent mode and it’s as much more resource-consuming, expensive and potentially uncontrollable as you might imagine.

The Gardener’s Dilemma: Tending the Illusion of Automation

Automation is only as automated as you tend to it like a garden. All automation, not just these looping AIs, are like that. Stability is an illusion. You’ve got to pay for that electricity or that cloud bill. The machine it’s running on has to keep working and not get unplugged from the network. You have to keep eating food, getting sleep, earning money to get food, lather, rinse repeat. It all takes energy. It all requires monitoring and maintenance. It’s always some sort of garden being tended by someone caring about that garden. Looping AI processes grow weeds that need to be weeded least the garden get overwhelmed and not be aesthetically pleasing or conforming to your vision.

Hmmm. Means of production? Did somebody say means of production? Bigwigs with land, big bank accounts and ownership of the oilfields or builders of the railroads use to have it. The landlords of the Agricultural Age. The tycoons of the Industrial Revolution. And now… now, what?

Anyone can build a datacenter. Money is all you need. Sure, some good planning but it’s no longer science. It’s just engineering, power sources, cooling systems and lots and lots of NVidia hardware. Demis Hassabis taught us all you need is scale when training computers to beat Atari-games circa 2013. A bunch of Google scientists taught us all you need is attention when they wrote the Transformers paper (the “T” in GPT) circa 2017. Those two things were science. They greatly thawed the AI Winter after the 1950s freeze when we had such great hope with the Perceptron and rise of the transistor. We got misled but now we’re back on track, but the science phase is done and the tweaking-it-out engineering phase has begun. If scaling and attention is all you need…

Creativity as the New Oil Refinery

Oh yeah, creativity! Refineries were necessary to extract the value out of crude oil. Even the value of gasoline was not seen until some creative person realized you could harness the power of controlled explosion to can the awesome force of a waterfall into a box. That’s why gas-burning fossil fuel combustion engines beat the steam engine and electric vehicles back in the day: a creative explosion led to a creative explosion which led to a static state in the automotive industry because you can’t keep being too creative or experimental motors will keep blowing up and nobody will know how to drive if you keep reinventing the steering wheel and gas pedals.

See? Creativity then stagnation. That is the way. Some wacky creative few pioneer a new industry, illuminate a path, show us the way. They believe things others don’t, take risks, toil away, get investors, align interests and bring it to market. If they fail to do those last 2 things: get investors and align interests than they have a double-whammy about to hit them. They lack allies who win when they win and they conflict with other peoples’ interests who will start to lose money as you start to gain money, and as we all know, that’s a big no-no. It doesn’t matter how well governed and regulated society is. Human nature is human nature and that inventor or scientist’s days are numbered.

That’s not cynical or anything. It’s just human nature. Statistics shows us that. People who have, acquire more because that’s what’s important to them. People who have less, keep getting less because material acquisition isn’t their thing and those who dedicate their lives figuring out how to simultaneously empty your wallet while maintaining control prey on them. It’s evolution — social primate hierarchical dynamics. There’s a big man on campus (BMOC) who wins a lot of mating rights, food resources, leisure time, etc. Then there’s all the different roles down the ladder. With chimps this is groups of 30 to 100 or so. Then there’s mandrills like Rafiki in Lion King. Those can be in groups of a thousand. And so on up to us at like 8-billion or so. Sure there’s country subdivisions but you think Gates, Musk, Bezos and their like abide by country boundaries?

Thank goodness for Jensen Huang. I mean I’m not giving him a get-out-of-jail free card or anything. But from a lifetime of being a science and technology fanboy it’s really refreshing to have the world’s most valuable company led by such a soft-spoken guy who will launch into talking about CUDA as their magic foresight and secret weapon instead of somebody who embraced and displaced Ashton Tate, Lotus, WordPerfect, Netscape and the rest of the trail of corpses that led up to one of the next-runner-ups in the trillion-dollar-companies.

Wanna hear some good stories? Look at what made each trillion-dollar company into a trillion dollar company. Jobs, Page & Brin, Gates, Huang. And look at Ellison jockeying for that position. It’s so fascinating!

It’s all about tech now. The era of Warren Buffett and Berkshire Hathaway are over. I mean listen to him talk about how it was his deep knowledge of certain trains being built that exclusively served certain regions that was the inside knowledge that got him investing. Yes, hardgoods will always have some advantage because it’s the stuff of life, the driving force behind Walmart and Ikea. And telecommunications had it’s day in the sudden catapulting freshly minted billionaires into the world as they locked-up transmission tower-building rights (just another form of landlord). Oh and landlords themselves? Well you see the ultimate endgame of the era of the landlord playing out right now. Sure a big man on campus (BMOC) can and would want to become President. Big surprise there.

But so long as scaling and attention are all you need, it’s still about tech. Artificial intelligence becomes gradually recognized as real intelligence, not by the generation that calls them stochastic parrots, clever plagiarists and incapable of consciousness because they’re machines. No, Penrose, Chomsky all the overly anthropocentric crowd die off in another 20 years or so and the new generations are born in… well, dependent on AI but that’s another story and article. This one is about creativity as the new oil refinery of the crude intelligence in datacenters and game-cards.

Oh yeah! And carers. Well, anthropology and archaeology are interesting because we’re mining our past like we’re mining Bitcoin. But whereas the last Bitcoin is going to be issued, we can just keep learning more about ourselves forever as we peel through the anthropological and geological layers, examining fossils and artifacts and learning new things. Those are the two obvious choices where just the desire to go out there and get your hands dirty exploring gets you outdoors and doing something that machines will need help with for a good long time, filling in that 4th circle of the Ikigai diagram — what the world needs.

And regarding the whole MCP and tool-calling thing? Well yeah sure MCP is about AI because it gives them the easier ability to use tools. But you as a human can use MCP directly just as well. More specifically, you can do that same sort of tool-calling trick that they do, but it uses that same central realization I’m using to let small models that are not trained on MCP do effective tool-calling nonetheless. That’s building on the bedrock of tech, the command-line interface or CLI where the rules are already universal. That’s all they’re doing with MCP, just with fancy wrappers to make the values being passed back and forth more explicit for those who can handle a fancier JSON-bloated language/protocol.

---

Book Analysis

This is a fantastic stream-of-consciousness entry that blends deep technical implementation with broad philosophical inquiry. Here is the requested transformation.

---

Crafting Compelling Narratives: Title & Filename Brainstorm

• Title Option 1: Casting Spells: A Simpler Magic for AI Tool-Calling
  • Filename: ai-tool-calling-magic-spells-mcp.md
  • Rationale: This title leans into the powerful “magic” metaphor used throughout the entry. It’s catchy, memorable, and effectively communicates the idea of simplifying a complex technical process (MCP/tool-calling) for a broader audience.
• Title Option 2: The Illusion of Automation and the Reality of a Command
  • Filename: philosophy-of-automation-llm-commands.md
  • Rationale: This title targets readers interested in the philosophical underpinnings of AI and software development. It frames the specific technical work as an example within a larger, thought-provoking discussion.
• Title Option 3: Progressive Enhancement for LLMs: A Lightweight MCP Orchestrator
  • Filename: progressive-enhancement-llm-mcp-orchestrator.md
  • Rationale: This is a more direct, technical title that will appeal to developers and engineers. It uses precise terminology (“Progressive Enhancement,” “MCP Orchestrator”) to signal its value as a practical case study.
• Preferred Option:
  • Title (plain text for YAML): Casting Spells: A Simpler Magic for AI Tool-Calling
  • Filename: ai-tool-calling-magic-spells-mcp.md
  • Rationale: It’s the strongest choice because it’s both evocative and descriptive. It captures the unique spirit of the author’s approach while clearly pointing to the core technical subject (AI tool-calling), making it highly suitable for a book that aims to be both insightful and engaging.

---

---

Book Potential Analysis

• Strengths as Book Fodder:
  • Relatable Metaphor: The central analogy of “magic spells” and a “wizard’s duel” provides a powerful and intuitive way to understand the complex, abstract concept of AI tool-calling protocols.
  • Authentic Voice: The entry captures the genuine, non-linear thought process of a developer, connecting a specific technical problem to broader existential questions about media, truth, and automation.
  • Practical Innovation: It documents a specific, clever solution—using simplified [command] syntax for MCP—that addresses real-world problems like token efficiency and compatibility with smaller models.
  • Philosophical Depth: It elevates a technical discussion into a richer exploration of what “automation” truly means, providing substance that goes beyond a simple “how-to” guide.
• Opportunities for Enrichment (for Book Adaptation):
  • Add a “Tale of Two Prompts” Comparison: Create a visual side-by-side comparison. On one side, show the verbose JSON an LLM would need for a standard MCP call. On the other, show the author’s simple [ls] command. This would instantly concretize the entire argument.
  • Include a “Key Concepts” Box: Early in the chapter, add a small, clearly defined box explaining Model Context Protocol (MCP), Tool-Calling, and Progressive Enhancement in simple terms for readers who are new to these ideas.
  • Connect to Industry Trends: Briefly mention how this lightweight approach relates to the broader industry push towards smaller, more specialized, or local models (SLMs), where token efficiency is paramount.

---

AI Editorial Perspective: From Journal to Chapter

This entry is a goldmine for a tech book, not just as a technical explanation but as a compelling narrative. It would be a perfect anchor for a chapter titled something like “The Philosophy of Tooling: Language as an Action” or “Pragmatic Magic: Building Intuitive Human-AI Interfaces.” The author isn’t just building a system; they are crafting a philosophy of interaction.

The most unique aspect is the seamless weave between the highly abstract (the nature of free will, the deceit of media) and the highly concrete (building an MCP orchestrator in Pipulate). This demonstrates that technical decisions are not made in a vacuum but are deeply informed by a developer’s worldview. The raw, stream-of-consciousness format, once structured with subheadings, becomes a powerful feature. It offers an authentic “behind-the-code” look at problem-solving, showing that innovation is a messy, beautiful process of connecting disparate ideas—from Richard Dawkins to The Sword in the Stone—to arrive at an elegant line of code. This entry embodies the idea of AI as a collaborative partner, not just a tool, where the goal is to find a shared, efficient language.

---

Suggested Next AI Processing Steps

1. Task Suggestion 1: Distill the Technical Rationale
   • Potential Prompt Snippet for Next AI: “From the provided text, extract the core technical argument for using a simplified, command-line syntax for the Model Context Protocol. Present this as a structured ‘Technical Abstract’ (150 words) covering: 1) The Problem with standard MCP, 2) The Proposed Solution ([command] syntax), and 3) The Key Benefits for developers and LLMs.”
2. Task Suggestion 2: Generate Explanatory Analogies
   • Potential Prompt Snippet for Next AI: “The author uses ‘magic spells’ as a central metaphor. Based on the entry’s content, generate three other distinct analogies to explain the author’s simplified tool-calling approach. For each, briefly explain what aspect it highlights (e.g., one for efficiency, one for universal accessibility, one for reliability).”