Articles
Practical guides on cutting the token usage of AI agents — Claude Code, Cursor, Codex, Copilot, Windsurf — without losing any capability.
LLM Gateway vs Token Optimizer: What Each Cuts
An LLM gateway like LiteLLM or Portkey controls where requests go and who pays for them. A token optimizer shrinks what the agent puts in each request. Different layers, different savings.
Paul IrollaRead article →THOL: Do Token Optimizers Really Save Tokens?
Every token optimizer quotes a big percentage. THOL checks what they actually do to the bill of a whole AI agent session. Here is how it works, what it found, and how to read it — no statistics degree needed.
Paul IrollaRead article →How to Reduce Devin Desktop Token Usage
Windsurf is now Devin Desktop, and its default agent changed with the name. What that means for your token bill, and the settings that cut it: ignore files, rules, MCP and modes.
Paul IrollaRead article →How to Reduce Droid Token Usage
Factory's Droid meters usage in rolling rate-limit windows. Here's how to see where they go, which settings cut them, and what no setting reaches.
Paul IrollaRead article →How to Reduce Grok CLI Token Usage
xAI's Grok Build gives you /context, /compact, a compaction threshold and token budgets on goals. Here's which settings cut the bill, and why its hooks can't trim tool output.
Paul IrollaRead article →How to Reduce Kilo Code Token Usage
Kilo Code bills every token at provider rates, so waste lands on your balance. The kilo.jsonc compaction knobs, model routing and read rules that actually move the bill.
Paul IrollaRead article →How to Reduce Qwen Code Token Usage
Qwen Code ships unusually good cost tooling: /context detail, /stats, an eager-tool allowlist, a compaction model. Here's which settings move the bill and what they leave untouched.
Paul IrollaRead article →Claude Code /compact: What It Costs and What Survives
A warm /compact costs cents, a cold one costs dollars on the same history. What compaction keeps, what it drops, and when /clear or /rewind is the cheaper move.
Paul IrollaRead article →Claude Code Hooks: Which Ones Add Tokens, Which Cut Them
Hooks can shrink what Claude reads or quietly add to every turn. The event you pick and the field you return decide which.
Paul IrollaRead article →Claude Code Subagents Token Usage: What They Really Cost
Every fresh Claude Code subagent pays a startup bill before it reads a file. How to read that bill per agent from the transcripts, and when delegation pays off.
Paul IrollaRead article →How to Choose a Token Optimization Tool
Four things drive an agent's token bill and each needs a different class of tool. Find which one is yours before comparing products, or you will buy the wrong category.
Paul IrollaRead article →Antigravity vs Gemini CLI: which costs less?
Two Google coding tools with opposite meters: one counts requests, the other measures how much work the agent did. That difference decides which one a heavy task punishes.
Paul IrollaRead article →When Does the Claude Weekly Limit Reset?
Anthropic publishes no universal weekly reset time, which is why every answer online disagrees. Here is what is actually documented, and where to read yours.
Paul IrollaRead article →opencode vs Kilo Code: which costs less?
Both are open-source agents you pay per token for, with no subscription absorbing waste. The difference is where they run and how you reach a model, and that decides which one leaks.
Paul IrollaRead article →How to Reduce Aider Token Usage
Aider sends a repo map with every request and, on some models, asks for whole files back. Those two defaults decide your bill, and both are one flag away.
Paul IrollaRead article →How to Reduce OpenClaw Token Usage
OpenClaw rebuilds its system prompt on every run, so your bootstrap files are billed again every turn. Here are the caps that control that, and the setting that reprices a whole request.
Paul IrollaRead article →opencode Tokens Per Second: How to See Them
opencode has no built-in tokens-per-second readout. Here are the three ways to get one today, and why throughput is the wrong number to optimise a bill against.
Paul IrollaRead article →tokensave Review: What Their 88% Covers
Three different projects are called tokensave. This reviews the one with 618 stars, reads the 88% figure it advertises for what that figure covers, and names the cost its tool count hides.
Paul IrollaRead article →caveman vs ccusage: the only pair you can settle in a week
caveman advertises around 65% off output tokens. ccusage tells you what share of your bill output tokens actually were. One of those numbers decides whether the other matters.
Paul IrollaRead article →caveman vs graphify: two ways to handle your docs
graphify ingests your documentation so the agent can query it. caveman rewrites the one document you re-send on every single turn. Both are about docs and they meet nowhere.
Paul IrollaRead article →caveman vs token-optimizer: output style or structure maps?
Both publish credible numbers, on opposite sides of the pipe. One is free for everyone; the other is free only if you are not a company. That decides more than the features do.
Paul IrollaRead article →caveman vs tokensave: speech or structure?
One changes how the agent talks, the other changes what it has to read. They are the two cheapest installs in the category and they solve nothing in common.
Paul IrollaRead article →ccusage vs graphify: did the graph even get used?
graphify is invoked by a slash command, which means it can be installed and never consulted. A meter is the difference between owning the tool and using it.
Paul IrollaRead article →ccusage vs token-optimizer: two dashboards, one independent
token-optimizer ships a dashboard richer than ccusage's. It is also reporting on its own work, which is exactly why you want the second, independent number.
Paul IrollaRead article →ccusage vs tokensave: measure it or index it?
tokensave publishes a retrieval benchmark on its own repo, not a session-cost figure on yours. ccusage is the only way you will get that second number, which makes this pairing less lopsided than it looks.
Paul IrollaRead article →codegraph vs caveman: index the repo or shorten the answers?
codegraph cuts the questions the agent has to ask. caveman cuts the words it uses to answer. There is one place where the two actually meet, and almost nobody uses it.
Paul IrollaRead article →codegraph vs ccusage: the tool and the instrument
codegraph publishes the strongest savings claim in this category. ccusage is how you find out whether that claim survives contact with your own repository.
Paul IrollaRead article →codegraph vs graphify: two graphs, two consultation models
Both build a graph of your project. One is always available over MCP, the other waits for a slash command, and that single difference decides more than their feature lists do.
Paul IrollaRead article →codegraph vs rtk: index the code or filter the shell?
Both are well-built, both are free, and they intercept completely different traffic. Which one pays off is decided by your session profile, not by their feature lists.
Paul IrollaRead article →codegraph vs token-optimizer: which file, or how big?
codegraph tells the agent which file to open. token-optimizer shrinks the file once it is open. They answer consecutive halves of the same question, which is unusual here.
Paul IrollaRead article →Codex Pricing: What It Actually Costs
Codex has no price of its own — it ships with every ChatGPT plan, and since April 2026 it bills token credits rather than messages. Here is what that changes for your bill.
Paul IrollaRead article →graphify vs token-optimizer: map the project or the file?
graphify builds one map of everything, up front. token-optimizer builds a small map of one file, on demand. Same idea at two scales, with very different maintenance.
Paul IrollaRead article →headroom vs ccusage: measure first, then compress
One of these tools reduces tokens and one only counts them. That sounds like an easy choice until you notice the counter is what tells you whether the compressor helped.
Paul IrollaRead article →headroom vs codegraph: compress or navigate?
One squeezes the conversation before every API call, the other stops the agent needing to read so much in the first place. Only one of those strategies survives a long session.
Paul IrollaRead article →headroom vs graphify: squeeze it or map it?
headroom compresses whatever reached the model. graphify tries to make sure less of it needed to. Both are Python, both are free, and they fail in completely different ways.
Paul IrollaRead article →headroom vs token-optimizer: two multi-layer bets
Both stack several techniques instead of picking one. The differences that decide it are the licence, the dashboard, and which of the two has been measured over a long session.
Paul IrollaRead article →headroom vs tokensave: compress or index the code?
A Python compressor that rewrites every API call against a Rust index that changes what the agent asks for. Same goal, opposite strategies, and very different risk profiles.
Paul IrollaRead article →How to Reduce Google Antigravity Token Usage
Antigravity meters work done, not prompts sent, and its quota refreshes on a five-hour clock. Here is what actually drains it and which built-in controls are worth using.
Paul IrollaRead article →rtk vs caveman: input side or output side?
rtk shrinks what your tools send the model. caveman shrinks what the model sends back. They attack opposite halves of the token bill, and only one half is usually the big one.
Paul IrollaRead article →rtk vs ccusage: two counters that disagree
rtk ships its own savings counter and ccusage reads your actual bill. When those two numbers disagree, the gap between them is the most useful thing either tool will tell you.
Paul IrollaRead article →rtk vs graphify: opposite ends of the same problem
One filters what comes back from your shell, the other indexes everything your project contains. They barely overlap, which makes picking between them a question about your own sessions.
Paul IrollaRead article →rtk vs token-optimizer: one mechanism or four?
rtk does one thing to your shell and reports it honestly. token-optimizer does four things and shows you a dashboard. One of the two also has a licence you must read.
Paul IrollaRead article →rtk vs tokensave: two Rust binaries, two channels
Both are compiled, both install in a minute, and both are free. That is where the resemblance stops: one owns your shell, the other owns your symbol graph.
Paul IrollaRead article →tokensave vs codegraph: which code index saves more?
Two MCP servers that index your codebase so the agent stops grepping. One publishes a session-cost benchmark, the other a retrieval one, and that difference decides more than the feature lists do.
Paul IrollaRead article →tokensave vs graphify: always on, or only when asked?
One is 80+ MCP tools charged on every turn whether you query them or not. The other costs nothing until invoked and gives nothing until invoked. That trade decides this page.
Paul IrollaRead article →tokensave vs token-optimizer: find it, or shrink it?
tokensave helps the agent find the right file across 50+ languages. token-optimizer makes that file 250 tokens instead of 180,000. One of them also needs a commercial licence.
Paul IrollaRead article →GitHub MCP Server: Setup and What It Costs
GitHub's MCP server is one of the heaviest you can leave connected. Here's how to install it, and how to load a tenth of it instead of all of it.
Paul IrollaRead article →headroom vs caveman: why the joke tool scored higher
One is a Python compression pipeline with an ML content detector. The other makes the model talk like a caveman. On the same benchmark, the joke placed fourth and the engineering placed last.
Paul IrollaRead article →How to Reduce Cline Token Usage
Cline is BYO-key, so every token it reads is billed straight to you — and by default it will happily read your whole project. Here's what drives that bill and which controls are worth turning on.
Paul IrollaRead article →How the Context Window Drives Your Token Bill
Your AI coding agent re-sends its entire context window on every turn, so the window size — not the answer length — is what actually sets your token bill.
Paul IrollaRead article →Prompt Engineering for AI Coding Agents
Prompt engineering still matters for coding agents — it's just no longer the expensive part. Here's what survives the move from chat to agent, and what quietly stopped mattering.
Paul IrollaRead article →Cursor Pricing: What You Actually Pay
Cursor Pro is $20 a month, and that number tells you almost nothing about what a heavy month costs. Here's where the meter runs and what changed when it moved to token billing.
Paul IrollaRead article →The Real Cost of Vibe Coding at Scale
Vibe coding feels free until you run it across a team for a quarter. Here's where the token bill actually comes from, the napkin math, and how to keep it from compounding.
Paul IrollaRead article →GitHub Copilot Pricing: What You Actually Pay
Copilot moved from counting requests to spending credits, and most comparisons still quote the old model. Here's what the plans cost now and where the meter actually runs.
Paul IrollaRead article →How to Reduce Windsurf Token Usage
Windsurf's Cascade agent burns tokens on indexed codebase retrieval, raw terminal output, MCP manifests and long Flows. Here's how to cut each one without making the model dumber.
Paul IrollaRead article →Token-Efficient MCP Tool Use
Token-efficient MCP tool use means paying for a server's schema only when a tool fires and compacting what it returns — not pre-loading manifests on every turn.
Paul IrollaRead article →LLM Cost: What Drives It and How to Cut It
LLM cost isn't set by the price per token — it's set by how many times you send the same tokens. Here's what actually drives the bill, and the levers that move it.
Paul IrollaRead article →Run Coding Agents Locally With Ollama
Running a coding agent on a local Ollama model cuts your API bill to zero — but only for the right jobs. Here's where local wins, where it doesn't, and how to split the work.
Paul IrollaRead article →MCP Servers: What They Are and What They Cost
An MCP server gives your AI agent tools it didn't ship with. It also bills you on every single turn, whether you call those tools or not — here's the mechanism and the maths.
Paul IrollaRead article →Claude Code MCP: How to Set It Up
Connecting an MCP server to Claude Code takes one command. The part nobody tells you is that every server you connect bills you on every single turn.
Paul IrollaRead article →Claude Code Pricing: What It Actually Costs
Claude Code ships with a Claude subscription and also runs on pay-as-you-go API billing. Here's what each route costs, and which one wins at your volume.
Paul IrollaRead article →Claude Code Skills: What They Are, What They Cost
Skills are markdown files Claude Code invokes on its own. Their descriptions sit in your context from the first message — here's what that costs you.
Paul IrollaRead article →When Does Claude's Limit Reset? 5-Hour and Weekly Times (2026)
Claude has two limits with two different reset clocks: a rolling 5-hour session window and a fixed weekly reset. Here's when each one comes back.
Paul IrollaRead article →Claude Pro vs Max: Is the Max Plan Worth It?
Max costs five times Pro and buys you five times the usage — not extra features. Here's how to tell which side of that line you're on before paying.
Paul IrollaRead article →Cursor vs Claude Code: Which One in 2026?
Cursor and Claude Code cost the same $20 a month and solve the same problem in opposite ways: one is an editor that gained an agent, the other an agent that never wanted one.
Paul IrollaRead article →headroom vs rtk: which one actually cuts your bill?
Both promise big token savings. On the open THOL benchmark, rtk broke even and headroom cost 53% more than running nothing — here's why, and what to do instead.
Paul IrollaRead article →OpenClaw: What It Is and What It Costs
OpenClaw is a self-hosted AI assistant you run on your own devices and message like a person. It's always on — which makes its cost profile unlike any session-based agent.
Paul IrollaRead article →How to Reduce opencode Token Usage
opencode is BYO-key, so every wasted token is billed straight to you. Here's what actually drives the bill and which built-in controls are worth turning on.
Paul IrollaRead article →Best Open-Source AI Coding Agents (2026)
A tested, criteria-driven ranking of the open-source AI coding agents that actually ship in 2026 — Aider, OpenHands, Cline, Goose, Continue, SWE-agent and Plandex — with honest limits and the real token cost of running each one.
Paul IrollaRead article →Stop an Overnight AI Agent Burning Your Budget
An agent left running overnight bills by the token, not the hour. Here's how to cap the spend before you wake up to a four-figure invoice.
Paul IrollaRead article →Context Engineering vs Prompt Engineering
Prompt engineering tunes one instruction. Context engineering controls the whole payload your coding agent re-reads every turn — which is where accuracy and cost actually live.
Paul IrollaRead article →Effective Context Engineering for AI Agents
Context engineering is the highest-leverage skill for running AI coding agents cheaply. Here's the mechanism, the techniques that actually move the bill, and the anti-patterns to kill.
Paul IrollaRead article →Agentic Engineering: The Discipline of Cheaper Agents
Agentic engineering is the discipline of designing what an AI coding agent reads, runs, and remembers per turn. Get it right and your agents get cheaper and more accurate at once.
Paul IrollaRead article →What Is Agentic Terminal Coding?
Agentic terminal coding is when an AI agent runs in your terminal and drives a tool-using loop — reading files, running commands, editing code — to finish a task on its own.
Paul IrollaRead article →Agentic Coding Best Practices for Sane Token Costs
The best agentic coding practices aren't about prompting tricks — they're about controlling what the agent reads. Here's how I keep token costs from spiraling.
Paul IrollaRead article →How to Reduce Tokens on Long Agent Sessions
Long agent sessions get expensive because every turn re-bills the whole transcript. Here's why cost grows quadratically, and the levers that actually flatten it.
Paul IrollaRead article →How to Measure AI Agent Token Usage
You can't cut what you don't measure. Here's how to actually quantify your AI coding agent's token usage — per call, per session, per dollar — instead of guessing.
Paul IrollaRead article →Lazy MCP Loading: Stop Paying for Idle Tools
Lazy MCP loading defers a server's tool manifest until you actually call a tool, so the per-turn overhead you pay on every message drops to near zero.
Paul IrollaRead article →Semantic Code Search vs Grepping the Repo
Grep finds every line that mentions a word; semantic code search finds the few that actually answer your question. For an AI agent paying by the token, that gap is the whole bill.
Paul IrollaRead article →Skeleton Compression: Read Files for Fewer Tokens
Skeleton compression hands an AI coding agent a file's structure — signatures, types, exports — instead of every line. Same understanding, a fraction of the tokens.
Paul IrollaRead article →Output Filtering: Trim Command Logs for Agents
Command logs are the silent token hog in agentic coding. Output filtering trims the noise before the model reads it, cutting cost without losing the signal you actually need.
Paul IrollaRead article →CLAUDE.md: How a Bloated File Inflates Your Bill
CLAUDE.md is re-read on every turn and lives in the cache prefix. Let it sprawl and you pay for it on every request, all session long. Here's the math, and how to keep it lean.
Paul IrollaRead article →Prompt Caching: How to Cut Your Input Bill
Prompt caching makes a stable prefix cost 10% or less on repeat reads instead of full price. Here's how to structure prompts so the cache actually hits, and where it stops helping.
Paul IrollaRead article →How to Reduce Cursor Token Usage
Cursor burns tokens on @-codebase context, indexed retrieval, MCP manifests and long agent threads. Here's how to cut each one without dumbing the model down.
Paul IrollaRead article →Claude Code: Subscription vs API Pricing
Claude Code runs on a Pro/Max subscription or on pay-as-you-go API billing. Here's the honest break-even math, and why token discipline changes the answer.
Paul IrollaRead article →Claude Usage Limits: How They Work
Claude usage limits aren't a hardware ceiling — they're a token budget. Here's how they actually work across plans and the API, and how to stop hitting them so early.
Paul IrollaRead article →Cut Claude Code Costs, Keep the Model
You don't have to drop from Opus to Haiku to cut your Claude Code bill. The cheaper move is to stop feeding the expensive model tokens it never needed.
Paul IrollaRead article →Claude Code Limits: How to Stay Under Them
Claude Code's limits aren't a wall you hit at random — they're a token budget you can spend slowly. Here's how the 5-hour and weekly caps work, and how to stay under them.
Paul IrollaRead article →How to Reduce Codex Token Usage
Codex bills you for eager file reads, raw command output, the MCP manifest and a growing transcript. Here's how to cut each one without losing quality.
Paul IrollaRead article →Agentic Coding: What It Is and Its Real Cost
Agentic coding is when an AI agent plans and executes multi-step coding tasks on its own. That autonomy is powerful — and it's why token costs can spiral fast.
Paul IrollaRead article →Best Claude Code Token Optimizers (2026)
Ranked roundup of every real Claude Code token optimizer — rtk, claude-context, codegraph, tokensave, token-optimizer, ccusage and Tokenade — with honest strengths and real limitations.
Paul IrollaRead article →Best Claude Code MCP Servers, Ranked by Real Cost
The best MCP servers for Claude Code, ranked by usefulness, real token cost and setup effort — including which ones quietly inflate every turn, and one that cuts that cost.
Paul IrollaRead article →What Is Vibe Coding?
Vibe coding is building software by describing intent to an AI and accepting what it produces — powerful for prototyping, expensive in tokens. Here's how it works and how to keep costs down.
Paul IrollaRead article →Best AI Coding Tools (2026)
Ranked: the 7 strongest AI coding tools in 2026 — Claude Code, Cursor, GitHub Copilot, Windsurf, Cline, Aider and Codex CLI — with real pricing, honest limitations, and the one layer that cuts the token bill across all of them.
Paul IrollaRead article →How to Reduce AI Coding Agent Token Usage
AI coding agents burn tokens by re-reading files, dumping directories and shipping verbose output every turn. Here are the levers that actually cut the bill — and how to apply them.
Paul IrollaRead article →How to Reduce Claude Code Token Usage
Claude Code burns tokens on eager file reads, unfiltered tool output, bloated MCP manifests and runaway transcripts. Here's how to cut each one without losing quality.
Paul IrollaRead article →Context Engineering for AI Coding Agents
Context engineering decides what your AI coding agent sees, in what form, and in what order. Get it right and you get better answers at a fraction of the token cost.
Paul IrollaRead article →