Saturday, July 18, 2026 · Claude users / solopreneurs meetup, Ahmedabad · ~22 min + Q&A. Deck: talks/52-percent-codemode (N or the notes button = condensed version of this script, on stage with you).
Walk on. Say nothing. Five to eight full seconds — people are scanning the QR; that's the plan, not a distraction.
Evening, everyone. That QR joins you into this talk from your phone — you'll raise hands and drop questions from there, so scan it now if you haven't. I'll wait.
Watch the join counter climb. When it slows:
Good. Keep that phone somewhere reachable. It becomes important again during the demo.
Tonight is about the bill your AI agent runs up — and specifically, the half of it you don't have to pay. That number on screen is the answer. The next twenty minutes are the question.
READ in sync with the screen, then pause.
What if I told you 52% is the floor — not the average, not the best case. The floor. If you pay for Claude, for the API, for any agent — this number is about your money. And 52 is what I measured in June — it has only gotten better since. Hold that thought for the receipt.
That's every room I've asked. By the end of this talk you'll know where every token goes — and how much of it you can stop paying.
Advance. The receipt counts up. Stay silent until the numbers settle. Don't read it — narrate the story in the lines.
This is a real receipt from my own MCPlex bundle. Three servers — PostHog, VaatChitra, and a memory server. Twenty-one tools between them. And look at the top line, because it's the most interesting one: PostHog recently collapsed their hundred-odd tools into a single exec tool. One tool. And its one schema is still six thousand six hundred tokens. Rent doesn't die. It moves.
The totals: eight thousand nine hundred tokens of JSON Schema becomes about a thousand tokens of compact TypeScript. When I measured this in June, the saving was 52 percent. Today the same receipt reads 88. That is why I said 52 is the floor.
And read the small print. This is the bill before anyone says a word. No question has been asked yet. This is purely the cost of showing the model its own toolbox.
Quick disclosure about who's standing here. I'm a solo consultant. I don't have a team. Claude is, effectively, my staff.
Beat.
Which makes this talk my staff's salary negotiation.
Deliver flat. Wait for the laugh — don't step on it.
And like any salary negotiation, step one is finding out what I'm actually paying for. Turns out there are two charges on the bill.
Charge number one: rent.
Every schema, for every tool, on every one of those servers, ships into the model's context on every single message. Whether the tool gets used or not. Seven servers is roughly sixty-seven thousand tokens — before you type a word. The amber is the schemas. The little green sliver at the end — that's you.
And here's the part that stings: it isn't a one-time security deposit. It ships again on every message. It's rent.
Charge number two: tax. But first — football. The final is tomorrow night, so this is legally required.
Right. Fifteen live, fifty highlights. Nearly everyone watched highlights. Of course you did. Ninety minutes of passes, compressed down to the goals. Hold onto that instinct — you already know how this talk ends.
Because here's what a live match actually is: tiki-taka. Hundreds of short passes. Now change one rule — every single pass has to go through the manager standing on the sideline. Player to manager, manager to player, every pass. The ball still moves, the game still works — but the manager touches everything. And the manager is the most expensive person on that pitch.
That's tool calling. Every call is a full inference round trip through the model, and the entire context — rent included — rides along on every trip. You're paying to watch every pass, live, at the manager's hourly rate.
So here's the shift, and you already voted for it five minutes ago: stop watching the match live. Watch the highlights. The model stops passing and starts writing. It writes a small program. The program calls the tools. Only the result comes back — the goals, without the ninety minutes.
Walk the traces left to right; don't read them.
Left side, the old way: fourteen round trips, and every byte of every intermediate result flows through context. Right side: one code block. What crosses back into the model is two hundred and twelve tokens. The answer.
Schemas load on demand. Intermediate data never touches context. Fourteen meetings become one email.
READ the quote with the screen — the engineers' line. Beat.
For the engineers in the room, the entire argument fits in one line: your model can code. Why are you bloating its context?
I needed this badly enough that I built a bundler for it — it's called MCPlex. You point your MCP servers at a bundle; the bundle gives you one endpoint that speaks codemode. That's genuinely all I'll say about it, because it's about to show up on its own.
Demo time.
Type the ask, narrating as you go:
Here's the ask, live. One: pull this week's Site Inspection API usage from PostHog. Two: check what the client actually said on our recorded calls — that's VaatChitra. Three: check what I noted about expectations — that's the memory server. And tell me: does the usage match what was promised?
While Claude works — this is your narration track, not dead air:
Things to watch while it thinks. These three services have never heard of each other. There is no integration between them. Nobody built one. Watch one code block join all three.
The exec beat — drop it mid-wait:
Fun fact while it works: PostHog recently collapsed their hundred tools into one exec meta-tool. Which means they've already conceded the rent argument. But exec still pays the tax — you ask it for tools, then for a schema, then to call. A full round trip each. Unless a program chains all three steps inside the sandbox — which is what's happening right now. PostHog fixed their rent. Codemode fixes their tax.
Result lands. Read the assessment aloud, briefly.
That analysis just crossed three services that don't know each other exists — on a program that was written, used, and discarded inside one request.
Call errors? Retry once, with a smile: "even the demo believes in retries." Wifi dead? → "Venue wifi is doing venue wifi things — here's the exact same run from last night." Play the fallback. Zero apology. Keep narrating over it.
READ with the screen. Slow down. Three beats. The quietest moment of the talk.
That integration was written.
Used once.
Thrown away.
Pause.
Nobody scoped it. Nobody maintains it. Nobody migrates it next year when one of those APIs changes. It cost about as much as the sentence describing it.
Now the business half — because cheap tokens are not actually the point.
You weren't wrong. You were pricing the wrong unit. Stop pricing tokens. Price decisions.
Walk the ladder bottom-up.
Most decisions in a workday are triage — a tiny model handles those, and it's the default. Everyday work: a small one. Real reasoning: mid-size. And frontier models are for high stakes only — the decisions where being wrong costs actual money. Run the ladder like this — any model, any agent stack, the shape is what matters — and most days the bill rounds to free.
So. Fifty-two percent. Remember — that was the floor. That's just the paperwork getting smaller. The sandbox savings and the collapsed round trips come on top, and nobody's charging you for those anymore.
READ the takeaway with the screen:
Here's the whole talk in one sentence: your agent doesn't need more tools. It needs a compiler and fewer meetings.
QR up.
If you'd rather see your own number than mine — that QR is MCPlex. Bundle your servers, and your bundle's page shows your percentage, live. Blog's at prashamhtrivedi.in. That's it from me — let's go to your questions.
Pull from the question queue — it's been filling since the demo. Read each question aloud before answering.
The talk is one open loop. "52%" appears in second one, gets a receipt in minute three, and is only fully explained by the hidden provenance slide — possibly never, if nobody asks. Everything between is a single argument in four movements: here's the bill → here's why it's high (rent + tax) → here's the fix (compiler) → here's what it means for your business (decisions, not tokens). If you feel lost mid-talk, find which movement you're in and land its one sentence.
Candid peer, verdict first, numbers concrete, zero hype. Never say "amazing," "game-changer," or "the future of." Say "six thousand six hundred tokens," not "a lot of tokens." Deliver the funny lines — salary negotiation, most expensive person on the pitch, fewer meetings — completely flat and let the room find them. Your credibility device is the receipt: every claim has a number, and the one number you can't fully defend on stage has a hidden slide behind the B key. That's the posture — a consultant who brought receipts.
Three waves. Quiet, deliberate open — the silence and the receipt do the work; resist filling them. Build through rent and tax — your most animated stretch. Peak at the demo — but the peak is their attention, not your volume. Then the hard drop: slide 10 is three sentences and near-silence, and it only lands if the demo before it had energy. Warm, brisk close. Loudest at the pitch diagram, quietest at "thrown away."
One join moment, four raise-hands (one a double), one question window. Two disciplines: never poll without using the result in your very next sentence, and remember only you see the counts — announce that early ("only I can see this, so be honest"), then always voice the number back. A silent poll is a wasted poll; the reveal is the moment. The crown jewel is the double poll on slide 6 — live minutes versus highlight reels. It reads as football banter but it's the thesis smuggled in, and slide 7 opens by cashing it: "you already voted for this."
Wifi dies: fallback recording, zero apology. A codemode call errors (known intermittent Workers I/O bug — fix before Saturday if possible): retry once with "even the demo believes in retries." Demo runs slow: narrate the architecture; thinking time is teaching time. Empty question queue: seed it — "someone always asks what the 52 actually measures" — and press B. Running long: compress slide 11 to the poll plus "price decisions, not tokens"; never cut the demo or slide 10. Hostile "this is just an ad": "Fair — MCPlex got forty-five seconds and a QR. The other twenty-one minutes work with any codemode setup."
Fix compact-types minification and redeploy, then verify the receipt numbers against the live bundle page. Record the fallback demo. Sanitize client-identifying PostHog data (RS Buildcon terms not closed). Load the one-line measurement answer. Scan all three QRs with your own phone. Rehearse thrice: once for words, once for silences, once fully live with the tool on a second device.