What you would be replacing
Caveman sits between your coding agent and the model provider as a local proxy, compresses logs, JSON, code, and search results with rules per content type, and keeps the original bytes recoverable through a retrieval tool. It also ships more experimental surfaces: rendering text to images for vision models, skill and memory-file compression, a browse driver, and cross-session memory. Any real alternative has to cover at least the first part, cutting what the agent sends without losing information, so that is the bar the options below are measured against.
The alternatives
1. Headroom: the closest like-for-like
Headroom works the same way: a local proxy that compresses agent context by content type before it reaches the model, with a retrieval tool so nothing is thrown away. The difference is packaging and scope. Headroom is a menu bar app for macOS, Windows, and Linux, signed and notarized on macOS, that sets up its own routing for Claude Code and ChatGPT Codex, starts with your machine, updates itself, and shows an activity dashboard of what it saved, including how much of your 5-hour and weekly usage windows remain. Caveman covers more agents and more experimental ground; Headroom covers Claude Code and ChatGPT Codex deeply, with experimental OpenCode and Grok Build connectors, and removes the assembly and maintenance work entirely. The full Headroom vs Caveman comparison lays out both sides of that trade.
2. Headroom CLI: free and open source
The Headroom CLI is the Apache-2.0 Python engine the app is built on. If what draws you to Caveman is running the whole stack yourself, this does the same job, free and under the same Apache-2.0 license Caveman moved to in 3.0. You give up the packaging (setup, auto-updates, the dashboard), which is exactly what the app vs CLI page says you are paying for.
3. Prompt discipline and built-in compaction: free, partial
The zero-install route: a be-concise instruction in your agent configuration, plus regular use of the agent's built-in context compaction. It costs nothing and helps at the margins, but it is not the same job. Instructions trim what the agent writes and tend to drift as sessions run long; they do nothing about the tool output, logs, and boilerplate the agent reads, which is where most of the token volume lives and where proxy-based tools earn their keep.
4. RTK: one category, done well
RTK rewrites noisy terminal commands into token-lean equivalents, so command output arrives already compact. It is not a full Caveman alternative, since it only touches one source of bloat, but it is worth knowing because it stacks with whichever proxy you pick rather than competing with it. Headroom ships it as a one-click add-on.
At a glance
| Alternative | Shape | Cost | Best when |
|---|---|---|---|
| Headroom | Menu bar app for macOS, Windows, and Linux, zero-config proxy | Paid, 7-day free trial | You live in Claude Code or ChatGPT Codex and want this handled, not administered. |
| Headroom CLI | Open-source Python proxy | Free | You want open source and are happy to install, configure, and update it yourself. |
| Prompt discipline | Config instructions + compaction | Free | You want zero install and accept it only trims output, not what the agent reads. |
| RTK | Terminal-command rewriter | Free | As a complement to any of the above, not a replacement for a proxy. |
How to choose
Every option here runs locally, so the test is the same one the vs page recommends: take a task you actually do, run it through each candidate, and compare your own numbers rather than anyone's published percentages - including ours. If your agents are Claude Code or ChatGPT Codex, Headroom is the alternative that asks the least of you. If you use agents beyond those two, or want the experimental surfaces, Caveman covers ground nothing else here does, and staying with it is a reasonable answer too.