The latest articles on DEV Community by Vikrant Shukla (@haltonlabs). https://dev.to/haltonlabs https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3920482%2F5bf7a95e-c7ed-4301-b55e-e60c1f995c2f.png https://dev.to/haltonlabs en Vikrant Shukla Fri, 08 May 2026 17:16:11 +0000 https://dev.to/haltonlabs/i-built-a-local-proxy-to-track-exact-llm-api-costs-per-project-1j82 https://dev.to/haltonlabs/i-built-a-local-proxy-to-track-exact-llm-api-costs-per-project-1j82 <p>The problem was simple: I run a small software studio that builds <br> client work heavily using Claude. Every project ended with the same <br> awkward conversation — "what did the AI actually cost?"</p> <p>Token estimates drift from the real bill. Nothing attributed costs <br> per project. I was either undercharging or handwaving, neither of <br> which builds client trust.</p> <p>So I built Halton Meter.</p> <h2> What it does </h2> <p>Halton Meter is a local mitmproxy-based daemon that intercepts <br> outbound LLM API traffic, attributes each request to a project, <br> computes exact cost from published pricing, and writes everything <br> to a local SQLite database. Nothing about how you call the API changes.<br> </p> <div class="highlight js-code-highlight"> <pre class="highlight shell"><code>pipx <span class="nb">install </span>halton-meter halton-meter init <span class="nt">--apps</span> </code></pre> </div> <p>Three processes come up:</p> <ul> <li>Edge listener on <code>127.0.0.1:8081</code> </li> <li>proxy interceptor on <code>127.0.0.1:8090</code> </li> <li>Loopback API on <code>127.0.0.1:8765</code> </li> </ul> <p>Every LLM API call you make — via the Anthropic SDK, OpenAI SDK, <br> raw HTTP, whatever — gets intercepted, tagged, costed, and logged.</p> <h2> Why a proxy and not an SDK wrapper? </h2> <p>SDK wrappers only catch calls you make directly. They miss ChatGPT, <br> Gemini Code Assist, and anything going through a tool you don't <br> control. A proxy captures everything on the wire without touching <br> your code.</p> <h2> Project attribution </h2> <p>The daemon attributes each request using a three-step chain:``</p> <p>Claude Code sessions, scripts, notebooks, and direct SDK calls all <br> get attributed correctly with zero code changes.</p> <h2> Terminal report </h2> <p><code>halton-meter report<br> </code><br> <a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F30j2q6zxes17avn2k4tj.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F30j2q6zxes17avn2k4tj.png" alt=" " width="800" height="471"></a></p> <p>Breakdown by project, model, and date. Numbers come directly from <br> the provider's published pricing — no estimates, no hidden margins.</p> <h2> What's supported </h2> <p>Six adapters across four providers: Claude, OpenAI, Gemini, and Grok. <br> Direct API surfaces and OAuth surfaces (ChatGPT, Gemini Code Assist) <br> both intercepted.</p> <h2> Local only </h2> <p>No cloud. No tracking. API keys never leave your machine. Everything <br> lives in <code>~/.halton-meter/db.sqlite</code>. The bundled dashboard is open <br> source (Apache 2.0) and runs locally.</p> <p>Docs and full architecture at <a href="https://haltonmeter.com" rel="noopener noreferrer">haltonmeter.com</a>.</p> <p>Happy to answer questions on the proxy architecture, the attribution <br> chain, or the cost calculation in the comments.</p> ai claude openai gemini