DEV Community

oVu X
oVu X

Posted on Originally published at caprock.dev

Caprock: See What Your Coding Agents Actually Cost

Caprock: See What Your Coding Agents Actually Cost

Coding agents feel free until the bill arrives. A coding agent is a program that reads, edits, and runs code from plain text tasks. The agent program costs nothing when it is open source. The model behind it charges for every call. Caprock shows that cost in one place. This article builds on the Caprock guide, updated 20 November 2026, and shows how it works, what it costs, and where it falls short.

What it is

Caprock is a free desktop app for macOS, Windows, and Linux. It is open source under Apache-2.0. It tracks every coding agent session on your machine. It records Claude Code, Codex, OpenCode, and more in one table. It shows usage and cost per repository, per branch, and per call. It also shows how close each plan sits to its limit.

Caprock reads session records from disk. It opens Claude Code and Codex transcripts, OpenCode and Hermes Agent databases, and DeepSeek Harness transcripts. It opens them read-only, so it changes nothing. Sessions from before you install it appear too.

All data stays on your machine in one SQLite file. SQLite is a small database that lives in a single file. Caprock needs no account, no cloud, and no telemetry.

What it costs, measured

The guide measures three agents on one machine. The numbers come from Caprock's own database, read on 8 October 2026. They are list-price models, not a bill. A list price is the price a provider lists for API calls. Do not read them as a ranking.

Agent Sessions Calls Total Median call
OpenCode 70 9,309 $156.28 $0.0023
Codex CLI 17 16,238 $621.30 $0.030
DeepSeek Harness 2 432 $0.88 $0.0018

OpenCode ran mostly DeepSeek V4 Pro and MiniMax M3. Codex CLI ran mostly GPT-5.6 and GPT-6 models. One model, gpt-5.6-sol, cost $503.46 of the $621.30 Codex total. Almost all input came from cache: 98.7% for OpenCode, 97.6% for Codex CLI, and 99.7% for DeepSeek Harness. Cache means the model reuses text it has already seen, at a lower price.

How to use it

Install Caprock first. On a Mac, run this command in a terminal:

brew install --cask dspv/tap/caprock-app
Enter fullscreen mode Exit fullscreen mode

The Mac app needs macOS 13 or later. On Windows, run this command in PowerShell:

irm https://caprock.dev/win.ps1 | iex
Enter fullscreen mode Exit fullscreen mode

Windows needs Windows 10 22H2 or later.

On Linux, run this command in a terminal:

curl -fsSL https://caprock.dev/linux.sh | sh
Enter fullscreen mode Exit fullscreen mode

Linux needs no sudo. Read the script before you run it.

Open Caprock after install. It finds your past sessions on its own. Pick a project, then start a new agent from the New agent button. It offers Codex, OpenCode, Gemini CLI, Grok Build, and Hermes Agent, beside Claude Code. Each agent opens in its own terminal tab for that project. Check the cost table first, before you start a long task.

Your first check takes one minute. Open the cost view and sort by repository. You will see what each project cost so far. Then open one session and read its calls. That first look tells you which agent and model cost the most.

Barriers, costs, and limits

Caprock itself is free. The models are not. You still pay your provider for every call, or you run a local model on your own hardware. A local model costs no API fee, but it needs a machine that can run it. On a flat plan, Caprock shows a list-price estimate beside your fee, never a bill.

Caprock does not cover every agent. It does not start or read Aider, Cline, Kilo Code, goose, or OpenHands. It starts Grok Build but does not read its cost. It reads Gemini CLI sessions only when Caprock starts them. Gemini CLI sessions you start outside Caprock stay outside its table.

The free build starts five agents. Custom agents need the paid Premium tier, and Premium reads no cost for them. The Windows installer is not code-signed yet. Windows SmartScreen asks once: click More info, then Run anyway. The session readers were tested on macOS. On Windows and Linux they build and their tests pass, but the guide has not run them there.

Treat the measured costs as one machine, not a rule. The three agents ran on different projects, in different months, on models with different prices. DeepSeek Harness has only two sessions in the sample. Your own costs will differ.

Verdict

Caprock solves one clear problem: you cannot manage what you cannot see. It puts every agent session and its cost in one local table. It asks for no account and sends no data out. That makes it a strong first install for anyone who runs more than one coding agent. Its gaps are clear too: some agents stay outside it, and some costs stay unread.

Install it if you already run Claude Code, Codex, or OpenCode. Skip it if you run only Cline or goose, because Caprock cannot see them. Start with the cost table, not with a new agent. The table alone may change which model you pick next.

Follow along

I test AI tools and build in public:

Top comments (0)