Add-ons, and why they're off by default
Headroom ships seven add-ons: RTK, Ponytail, Caveman, MarkItDown, Serena, Codebase Memory, and Context7. Some of the biggest token savings in the app sit behind their switches. So why would a token-saving app leave its biggest savers off? Because of a line we won't cross without you, and once you see the line, choosing your add-ons takes about a minute.
The line: the core proxy never changes behavior
Headroom's core compresses what your agent reads (tool output, logs, file dumps) while leaving what your agent does untouched. Same commands, same answers, same code. Add-ons cross that line on purpose: each one changes how your agent behaves in exchange for bigger savings. That's a trade you should choose, not one we should sneak in. So they're opt-in, one click each, in the app's Add-ons view.
RTK: quieter terminals
RTK wraps common shell commands (git, cargo, npm, test runners) and strips the noise before your agent ever sees it: passing-test spam, progress bars, boilerplate. Terminal output shrinks 60-90%. The behavior change: your agent sees summarized command output instead of raw output. See Headroom vs RTK for why they stack rather than compete.
Ponytail: leaner code out
Ponytail is a skill that steers your agent toward the laziest solution that actually works: standard library before dependencies, one line before fifty, no speculative scaffolding. It cuts output tokens, often the most expensive ones, and, many users find, produces better code. The behavior change is the point: your agent writes differently with it on. Details in Headroom vs Ponytail vs RTK.
Caveman: terser replies
Caveman is a plugin that makes your agent reply in terse caveman-speak, cutting output tokens while keeping code, commands, and errors exact. It pairs naturally with Ponytail: Caveman trims the words, Ponytail trims the code, and they never touch the same tokens. The behavior change: your agent's prose gets blunt. It installs into Claude Code and Codex directly, so it needs their CLIs and Node.js on PATH. See Headroom vs Caveman for how the plugin relates to the wider Caveman toolkit.
MarkItDown: documents your agent can afford to read
Agents can't read PDFs and Office files directly, and workarounds burn tokens. MarkItDown converts .docx, .xlsx, .pptx, and PDFs into clean Markdown locally, so your agent reads a tight text version instead. The behavior change: file reads on those formats route through the converter.
Serena: read one function, not the whole file
Serena is an MCP server that gives your agent symbol-level code tools (find a class, read a single function, trace references) so it reads the piece it needs instead of whole files. It saves most in large repos. The behavior change: your agent navigates code through Serena's tools, and those tool definitions ride along on every request, so small projects may not come out ahead.
Codebase Memory: a persistent map of your repo
Codebase Memory is an MCP server that indexes your codebase into a knowledge graph of call chains, classes, and routes, so structure questions ("what calls this?", "where are the API endpoints?") are answered from the graph instead of by re-reading files. It complements Serena: a pre-built map versus live symbol tools. The behavior change: your agent consults the index before it explores.
Context7: current library docs
Context7 is an MCP server that fetches current, version-specific documentation for the libraries you use, so your agent stops burning tokens on guessed or outdated APIs. The behavior change: your agent looks up real docs instead of trusting its training data. Requires Node.js 20.18.1 or newer on PATH.
Which should you enable?
If your sessions run lots of builds and tests, start with RTK. If your bills are output-heavy, add Ponytail for leaner code and Caveman for leaner prose. If your projects involve reading documents, MarkItDown pays for itself the first time. In large codebases, try Serena or Codebase Memory; if your work leans on fast-moving libraries, Context7 keeps the API guessing out of your bill. All seven coexist with the core proxy and with each other, and everything is reversible with the same single click.
See it on your own traffic: install Headroom and run a normal session.