52%

follow along on your phonetools.prashamhtrivedi.app · join

What if I told you 52% is the floor, not the average?

The receipt

MCPLEX · TOKEN RECEIPT

bundle: posthog · vaatchitra · memory

POSTHOG · 1 tool0

VAATCHITRA · 6 tools0

MEMORY · 14 tools0

UPFRONT CTX (JSON)0

CODEMODE (TS)0

YOU SAVE0%

* before the first question is asked

PostHog already collapsed ~100 tools into one — and its one schema is still 6,656 tokens.

Rent doesn't die. It moves. 52% was June's floor — today the receipt reads 88.

Who's asking

I'm a solo consultant. Claude is effectively my staff.

This talk is my staff's salary negotiation.

prashamhtrivedi.inthe consultant in question

Why tool calling is expensive · 1 of 2

Rent: schemas ship on every message

github

posthog

linear

slack

drive

memory

calendar

you

■ 7 servers ≈ 67,000 tokens of schema — before you type a word ■ your actual question and it ships again on every message

Why tool calling is expensive · 2 of 2

Tax: every call is a full round trip

You've all seen a World Cup match. Tiki-taka — hundreds of passes. Now route every pass through the manager on the sideline. That's tool calling: each pass a full inference round trip, rent riding along.

every pass, via the manager

The shift

The agent writes a program instead

Before · one round trip per call

> call posthog.query        ↩ 2,340 tok
> call posthog.query        ↩ 1,872 tok
> call memory.find          ↩ 3,405 tok
> call memory.get           ↩ 2,911 tok
  … ×14 — every byte through context

After · one code block

> code {
    const usage = posthog.query(…)
    const notes = memory.find(…)
    return assess(usage, notes)
  }                          ↩ 212 tok

Your model can code. Why bloat the context?

MCPlex is the bundler I built out of necessity: point your servers at one bundle, and the bundle speaks codemode.

/prasham/bundles/analytics/mcp-code → 1 tool: code

Live · one question, three strangers

Demo time!!!

That integration was

Written.
Used once.
Thrown away.

Nobody scoped it, nobody maintains it, nobody will migrate it next year.

For the solopreneurs

Cost per decision, not cost per token

Tinytriage · the default

Smalleveryday work

Mid-sizereal reasoning

Frontierhigh-stakes only

Any model, any agent. Escalate deliberately, and most days the bill rounds to free.

52% was the floor

Your agent doesn't need more tools. It needs a compiler and fewer meetings.

mcplex.prashamhtrivedi.appbundle yours, see your own %

52% of what, exactly?

What was measured

Upfront tool-definition context. Every tool serialized as JSON Schema (name, description, inputSchema) vs the compact TypeScript declarations codemode sends instead. Tokens estimated at ~4 chars per token.

On what workload

My live MCPlex bundle — the same one in the demo: PostHog (1 exec tool, a 6,656-token schema), VaatChitra (6 tools), Memory (14 tools). Computed per bundle and shown on its public bundle page, live: your bundle shows your number, not a benchmark. 52% was the June measurement; after PostHog collapsed their tools, today's reads 88%.

Why it's the floor

The 52% counts only the upfront context. The bigger runtime wins — results staying in the sandbox (return is all that crosses back) and N round trips collapsing to one — aren't in the number at all.

“Why not just fewer tools?”

It's not about tool count. Once tools live behind one code tool, any two of them can compose — PostHog × Memory today, anything × anything tomorrow. Composition is the leverage; pruning isn't.