Beyond token savings
"I'm on a subscription, so I don't pay per token, so what does compression actually buy me?" Fair question, and the first answer is hiding in your plan's fine print: your limits are token-denominated even when your bill isn't. That's one of four things Headroom does that have nothing to do with an invoice. Here's the whole list.
Subscribers: more work per window
Claude and ChatGPT plans meter you in tokens per 5-hour window and per week; see Claude Code's limits and Codex's limits. Every token Headroom strips from tool output and logs is a token your limit never sees, which stretches the same plan further: longer sessions before the wall, fewer Wednesday lockouts, less pressure to upgrade a tier. As one of our beta testers put it: Headroom "might extend our usage beyond the quantity of tokens we would normally be able to use." That's the subscription version of saving money.
API users: the direct version
On the API you pay per token, so compression is a line item on your invoice. Our dashboard prices what was actually avoided, and deliberately never counts your provider's prompt-cache discount as our work. The exact definition is in how savings are measured.
An agent that learns your machine
Compression saves tokens this session; auto-learning saves them every future session. Headroom notices what your agent keeps getting wrong (broken commands, wrong paths, ignored preferences) and writes the confirmed lessons into your agent's context files. Fewer failed round-trips is both a cost win and a quality win: the agent spends its context on your problem, not on rediscovering your environment.
Visibility you didn't have
The dashboard shows where your usage actually goes: sessions, savings by day and hour, your position inside Claude's and Codex's usage windows, and which optimization layers did the work. Most people's first surprise isn't the savings; it's seeing what was burning their window in the first place.
One place for the whole toolkit
Token efficiency is an ecosystem: RTK for terminals, Ponytail and Caveman for output, MarkItDown for documents, Serena and Codebase Memory for code navigation, Context7 for current library docs. Wiring those up by hand means proxies, shell hooks, plugins, and MCP config. Headroom packages all seven as one-click add-ons behind a single app that keeps itself updated.
See it on your own traffic: install Headroom and run a normal session.