Markdown version of https://extraheadroom.com/headroom-alternatives # Headroom alternatives: what else cuts agent tokens Headroom compresses what Claude Code, ChatGPT Codex and other coding agents send to the model, through a proxy on your machine. These are the other tools that do that job, or part of it, and what each one does differently. We make Headroom (the open-source Headroom CLI and the Extra Headroom app built on it), so this page also says where the benchmarks that rank Headroom last come from. ## Quick answer The closest alternative is Caveman, an open-source toolkit that added its own compression proxy in version 4 and covers more agents. context-mode takes a different route: hooks keep large tool output out of the context instead of compressing it on the way out. SuperCompress does the work on its own servers rather than on your machine. If you run the Headroom CLI and mainly want less upkeep, the Extra Headroom app is the same engine, set up and updated for you. ## At a glance | Tool | How it works | License | Price | | --- | --- | --- | --- | | Caveman | Local proxy plus terse-output skill | Apache-2.0 | Free | | context-mode | Hooks and a local sandbox | Elastic-2.0 (source-available) | Free plugin, hosted gateway from $20 a month | | Context Gateway | Local proxy using small compression models | Apache-2.0 | Gateway free, API on enterprise terms | | SuperCompress | Hosted model picks the relevant lines | MIT client, hosted engine | 1M tokens a month free, then $0.10 per 1M | | lean-ctx | Shell hooks, cached reads and a proxy | Apache-2.0 engine | Free, Pro $9 a month | | Tokenade | Local binary between agent and model | Apache-2.0 | Free | | pxpipe | Renders old context as images | MIT | Free | | Built-in compaction | The agent summarizes its own history | Part of the agent | Free | ## The alternatives ### 1. Caveman: the closest open-source match [Caveman](https://github.com/JuliusBrussee/caveman) began as a skill that makes the agent reply in fewer words. Version 4.0 (October 2026) added a proxy that shrinks logs, JSON, CSV, YAML and test output and keeps the originals on disk, which is the same core job as Headroom. It also routes requests between models through a free Caveman Cloud account. It covers Claude Code, Codex, Gemini CLI, OpenCode and more, and it is Apache-2.0 throughout since 3.0. Install it with `npx @caveman-ai/cli@latest`. Maker's claim: 33.2% fewer input tokens, over 54 runs in August 2026 on an earlier engine. The [Headroom vs Caveman comparison](/headroom-vs-caveman) goes deeper. ### 2. context-mode: keep tool output out of the window [context-mode](https://github.com/mksglu/context-mode) does not compress requests. Its hooks run large tool output in a local sandbox, index it, and give the agent only the parts it asks for. It supports Claude Code, Codex, Gemini CLI, OpenCode, Copilot CLI and more. The plugin is free; it is source-available under the Elastic License 2.0, not open source. In Claude Code, run `claude plugin marketplace add mksglu/context-mode`, then `claude plugin install context-mode@context-mode`. Maker's claim: 99.4% of output bytes saved, on runs with more than 10 KB of output. That counts bytes kept out of the window, not billed tokens. ### 3. Context Gateway: compression by small models [Context Gateway](https://github.com/Compresr-ai/Context-Gateway), from the YC company Compresr, is a local proxy that compresses tool output with small models, guided by what the tool call was for. The agent can expand any result back to the original. It also prepares compaction in the background before the window fills. The gateway is Apache-2.0 and supports Claude Code and Cursor; the compression API is sold on enterprise terms, with keys issued after a call. Its last release was in March 2026. Maker's claim: up to 90% fewer tokens, with no method published. ### 4. SuperCompress: compression as a hosted service [SuperCompress](https://supercompress.dev) sends your agent's context to a hosted model that keeps only the original lines relevant to the current question. The client is MIT-licensed and installs with `npx -y supercompress-ai setup`. The first 1M tokens a month are free, then $0.10 per 1M. The difference that matters most: your context leaves your machine to be compressed, where Headroom and the other tools here run locally. Its own benchmarks page puts its headline figures "on hold pending re-verification". ### 5. lean-ctx: hooks, cached reads and a proxy [lean-ctx](https://github.com/yvgude/lean-ctx) is a Rust binary that compresses command output through shell hooks, serves cached file reads over MCP, and can also run as a proxy. The engine is Apache-2.0, with paid Pro ($9 a month) and Team plans. Install it with `curl -fsSL https://leanctx.com/install.sh | sh`, then `lean-ctx wrap claude`. Its README withdrew earlier savings figures and says results range from a clear win to break-even. ### 6. Tokenade: a small local binary [Tokenade](https://github.com/pi-infected/tokenade) is a Rust binary that sits between the agent and the model. It folds command output, skips repeated file reads and compacts stale context. It is free, Apache-2.0, and supports Claude Code, Codex, Cursor, OpenCode and more. Install it with `npm install -g @tokenade/cli && tokenade install`. Its author also runs THOL, the benchmark below. ### 7. pxpipe: context as images [pxpipe](https://github.com/teamchong/pxpipe) renders the system prompt, tool docs and older history as images, which the model bills by size rather than by character. It is MIT and free. Its maker reports 59 to 70% lower bills on a June 2026 model, and says the Claude models released since then "regressed sharply on dense pixel reading". Treat it as an experiment. ### 8. For OpenCode only: DCP and Magic Context Two OpenCode plugins prune history inside the agent. [DCP](https://github.com/Tarquinen/opencode-dynamic-context-pruning) (AGPL-3.0) swaps stale parts of the conversation for summaries and drops repeated tool calls; its maker says development has slowed. [Magic Context](https://github.com/cortexkit/magic-context) (MIT) compresses history in the background so compaction never pauses the session. Headroom has an [OpenCode connector](/headroom-opencode) too, still marked experimental. ### 9. Built-in compaction: free, and partial Claude Code and Codex both summarize their own history when the window fills, and `/compact` does it on demand. It costs nothing to install. It shrinks the history, not the tool output arriving on each new turn, and the summary request is itself a large one. The [free built-ins](/best-claude-code-token-saving-tools#built-in) go further. ## About the benchmarks that rank Headroom last Several pages, some by the makers of the tools above, report that Headroom raised costs. Here is what each one tested: - **THOL** , run by Tokenade's author, who notes it "is not independent": Headroom 0.27.0 in June 2026, 52.8% more expensive than no tool on long tasks. An earlier run of the same benchmark on the same version scored level with no tool. - **Marmelab** , August 2026: one task, three runs. 32% more expensive in one mode, 4% cheaper in the other, which they called noise. - **Caveman's comparison page** : Headroom 0.33.0 in August 2026, 6.7% saved. The cause these pages give is the prompt cache. When compression changes a turn the provider has already cached, the cache discount on that turn is lost, and on long sessions that cost more than compression saved. That was a fair finding. Since engine version 0.38.0, Headroom resends the part of a conversation the provider has already cached byte for byte and compresses only what is new. The current release is 0.42.0. Every published run we found tested 0.36.4 or older, and none tested the Extra Headroom app. So don't take our word either. Run a week of your normal work with and without Headroom, and compare the cost per session in Claude Code's `/usage` or in [ccusage](https://github.com/ccusage/ccusage). ## How to choose - **You want it handled in Claude Code or ChatGPT Codex:** the Extra Headroom app sets itself up, updates itself and shows what it saved. - **You want open source across many agents:** Caveman, or the free [Headroom CLI](/headroom-desktop-vs-cli). - **You would rather not run a proxy:** context-mode. - **Your context must stay on your machine:** anything here except SuperCompress. - **You only use OpenCode:** Magic Context. Run one compressor at a time. Two on the same request fight over the same tokens, and you can no longer tell which one saved what. Tools that cut different tokens do stack; the [best Claude Code token-saving tools](/best-claude-code-token-saving-tools) list shows which. ## How we checked Licenses, prices, install commands and supported agents come from each project's own repository or site, checked on November 5, 2026. Savings figures are each maker's own, measured on different tasks, so they do not compare with each other.