Notes on running agents without losing track of the bill.

Auto-retrying to the cloud when your local model gives a weak answer

Local models save money until the day they quietly return garbage and nothing catches it.

How to actually raise your cache hit rate with a coding agent

Cache hits are a property of how your session is shaped, not a setting you turn on.

ccusage alternatives that show live spend, not just history

A log reader tells you what a run cost. Something in the request path can change what the next one costs.

Does Claude Code's context management actually save money

Compaction buys you a longer session. The money question is what it does to your prompt cache.

How much does Claude Code actually cost per day

Token math, a worked day, and the three numbers that decide what you actually pay.

Claude Code latency vs cost: is the fast model worth the extra spend

The answer depends on whether a human is sitting there waiting, and almost nothing else.

Claude Code Pro vs Max vs API billing: which is cheaper for you

Stop guessing at your tier. Measure what you actually spend, then pick.

Why Claude Code subagents multiply your cost so fast

Fan-out looks like parallelism. On the invoice it looks like running the same job several times over.

Tracking Codex CLI costs across projects

Codex writes enough to its session files to rebuild a cost report. Here is how to do it, and where that method quietly falls apart.

Cursor vs Claude Code: which one burns more tokens

One tool shows you the meter. The other one turns it off. Neither of those is the same as being cheap.

Deduplicating simultaneous identical LLM requests in an agent workflow

Your cache only helps after the first response lands. This is about the gap before that.

Does compressing your context break prefix caching savings

Shortening a prompt looks like a saving until you check what it did to your cache hit rate.

Exact-match vs semantic LLM caching: which one is safe for agents

One replays a byte-identical request. The other guesses that two requests mean the same thing. Only one of those is safe to leave on by default.

Why more dev tools are dropping email/password for Google or GitHub sign-in

Fewer passwords to leak, fewer forms to abuse, and a login flow that already knows who a developer is.

Installing a local HTTPS proxy certificate on macOS for dev tools

The five-minute keychain step that every local proxy setup guide skips.

How to keep proprietary code off cloud LLM servers entirely

Three ways to stop coding agents from shipping your source to someone else's GPU, and where each one actually breaks down.

Langfuse vs a local request ledger: do you need a hosted observability stack

Tracing platforms answer questions about a product you ship. A local ledger answers questions about the bill you pay.

Pointing your coding agent at LM Studio instead of the cloud

The server part takes five minutes. The part where the agent still finishes the task is the interesting one.

When is a local model good enough for coding, and when isn't it

Skip the leaderboard. Here is the actual test I use before a request goes to Ollama instead of a frontier model.

Best local models for coding on an M-series Mac in 2026

Memory bandwidth, prefill speed and tool-calling reliability decide this, and the answer is very different at 16GB than at 128GB.

What you give up (and gain) using a macOS-only private beta dev tool

Platform lock-in is a real cost. So is pretending it isn't.

Model tiering: routing to a cheap model first, only escalating on failure

Most agent turns don't need your most expensive model. Route cheap first, promote on evidence of failure.

Running Claude Code, Codex, and Cursor together: what it costs

Three subscriptions, three terminals, zero combined view of where the money goes.

Why a local-only proxy is different from a hosted AI gateway

The trust model changes when there is no server to trust.

Using Ollama with Claude Code and Cursor to cut cloud costs

Local models are good at a narrower slice of agent work than people claim. The trick is knowing which slice.

One proxy for Claude Code, Codex, and Cursor: is it worth setting up

The honest version: worth it once you run more than one agent, and only if the proxy is local.

Attributing AI agent spend to the process that actually made the call

The invoice tells you what model. It never tells you which terminal tab.

Prompt caching vs response caching: what actually saves money

One is a discount on a request that still happens. The other stops the request from happening at all.

How to reduce Claude Code token usage without losing context

The tactics that actually move the number, and the ones that quietly make things worse.

Can you run an AI coding agent fully offline

The honest answer, not the marketing one.

Why semantic caching breaks agent tool calls (and how to avoid it)

A vector similarity hit above your threshold is still a guess. Guesses inside a tool call chain corrupt the whole run.

What actually happens to your code when you send it to a cloud LLM

The request path, the retention window, the training question, and the parts your agent sends without asking you.

How to set a hard spend limit on an AI coding agent

Dashboards tell you what you spent yesterday. A cap stops the bleeding today.

Setting a similarity floor so semantic cache never corrupts an agent run

The threshold number matters less than what you refuse to match at all.

How do you detect a weak LLM answer well enough to auto-retry it

Confidence scores lie. Here is what actually works.

Why AI coding agents burn through API credits so fast

It isn't one big prompt. It's hundreds of small ones, and almost nothing is watching them.

Why is Claude Code so expensive all of a sudden

Your usage did not change. Your token economics did.