What each piece actually is
| Piece | What it is | How you turn it on |
|---|---|---|
| Headroom | A local proxy plus a macOS menu bar app. Not a plugin, not a skill. | Download and open once. It then applies to every session. |
| Ponytail | A genuine Claude Code and Codex skill that makes the agent write less code. | One-click add-on, or install by hand. |
| RTK | A CLI wrapper that trims noisy terminal command output. | One-click add-on in the Headroom app. |
| MarkItDown | A converter that turns PDFs and Office files into compact Markdown. | One-click add-on in the Headroom app. |
Why a proxy instead of a plugin
This is a deliberate design choice, not a gap in the roadmap. A plugin only sees what the agent chooses to hand it, and only in the sessions where you remembered to turn it on. A local proxy sits on the request path, so it sees the request the way the model does: the whole assembled context, every time.
That buys three things a plugin cannot. Coverage is total, including sub-agent runs, which are exactly where token costs spike because they lose the parent session's prompt cache. There is no per-session setup, so you cannot forget to enable it. And compression can be reversible: because the original content stays on your machine, Headroom can attach a retrieval tool that lets the model pull back the full detail whenever it needs it.
If you came looking for a skill
You probably want Ponytail. It is a real Claude Code skill (and it works with Codex), it has slash commands and intensity levels, and it attacks a side of the bill that Headroom's compression does not touch: the code the agent writes back. The install guide covers both the one-click and manual paths.
Is there an MCP server?
The open-source Headroom CLI can register itself as an MCP server, and the app sets up an MCP entry for retrieval so the model can pull back compressed content on demand. That is a supporting mechanism rather than the main event: the compression itself happens in the proxy, not through MCP tool calls, which is why it works without the agent cooperating.
What this means in practice
Install the app once and every Claude Code and Codex session is optimized from then on, including after a restart, with no plugin list to curate and no command to remember. Then toggle on the add-ons you want from one screen. If you would rather run the engine yourself from the terminal, the open-source CLI is free and does the same compression — see app vs CLI for the trade-offs.