Privacy-first analytics pixel

For AI agents: this site's index is at /llms.txt, and each docs, guide and blog page comes as Markdown at its URL plus .md.

Headroom logo Headroom

Add-ons

Headroom ships eight optional add-ons that can contribute additional savings through shorter command output, more targeted reads, or shorter generated responses.

Enable an add-on

  1. Open Add-ons in Headroom.
  2. Choose an add-on for your task and use its install or enable control.
  3. Wait for installation to finish and resolve any prerequisite shown by the app.
  4. Start a new agent session so it can load the updated tools or instructions.

Add-ons are optional because they change the agent's tools or instructions and may require additional software. Core compression works without them. Enable one at a time to evaluate its effect on your workflow.

RTK

Summarizes supported shell-command output before the agent reads it.

Use for: Builds, tests, git commands, and other tasks with repetitive terminal output.

Considerations: The agent receives filtered output. Check the full command output when diagnosing an error that needs more detail.

Ponytail

Adds instructions that favor small implementations, standard libraries, and fewer unnecessary dependencies.

Use for: Coding tasks where the agent generates more code or abstractions than the task requires.

Considerations: Changes how the agent approaches implementation. Review the resulting code and run the relevant checks. It can also skip optional steps in non-coding work, so turn it off for that. To turn it off in one Claude Code project, add "enabledPlugins": {"ponytail@ponytail": false} to that project's .claude/settings.json. Installation requires the agent CLI, Node.js and Git on PATH.

Levels, commands and the manual install are in how to install and use Ponytail in Claude Code.

Caveman

Adds instructions for shorter natural-language replies while preserving code, commands, and error text.

Use for: Sessions where explanations and progress messages account for substantial output.

Considerations: Replies become terse. Disable it when you need detailed explanations or onboarding help. Installation requires the agent CLI, Node.js and Git on PATH.

Chisle

Keeps replies terse and code minimal, and trims long command and web output.

Use for: One add-on that covers what Ponytail and Caveman do.

Considerations: Use it instead of Ponytail and Caveman, not alongside them: together they give the agent overlapping instructions in every session. Installs into Claude Code and Codex. Installation requires the agent CLI, Node.js and Git on PATH.

MarkItDown

Converts PDF and Office documents into Markdown for the agent to read.

Use for: Summarizing reports or extracting requirements from Word documents, presentations, and spreadsheets.

Considerations: Check the extracted text against the original when layout, images, or tables matter. Conversion does not guarantee that every document element is preserved. It converts files the agent reads from your project folder, not files attached to a chat.

Serena

Provides tools for finding symbols, reading individual functions or classes, and tracing references.

Use for: Navigating a large repository or investigating a change limited to a few symbols.

Considerations: Adds tools to the agent through MCP (Model Context Protocol). Those tools also add context; savings depend on whether they replace larger file reads.

Codebase Memory

Indexes repository structure in a persistent knowledge graph, including call chains, classes, and routes.

Use for: Repeated questions about how a codebase is organized or which components depend on each other.

Considerations: After installation, restart the agent session and ask it to index the repository before querying it. Indexes are stored in Headroom's app data. Confirm relevant details against the current source when the repository has changed. Serena provides live symbol tools; Codebase Memory provides indexed relationships.

Context7

Provides library documentation for the version you are using.

Use for: Implementing or upgrading an integration when the agent needs accurate API examples.

Considerations: Documentation lookup adds a tool call and returned content. Specify the library version in your request. Claude Code, Codex and OpenCode use Context7's hosted server, so your library lookups go to Context7 and nothing extra runs on your computer. Grok Build runs it locally, which needs Node.js 20.18.1 or newer on PATH. Context7's free tier allows 1,000 lookups a month, then 20 a day, as of November 2026.

Evaluate and disable add-ons

Run a representative task, review the result, and check Activity for the optimization records available. Savings vary with the task; an installed add-on does not necessarily run on every request.

Disable keeps the add-on installed for later use; Uninstall removes its managed installation or registration. For MCP add-ons, disabling removes the managed server registration. Restart agent sessions after either change. Headroom preserves MCP entries you configured yourself. For worked steps, see tutorials and use cases. The dashboard's Output indicator measures Headroom's output-shaping pipeline, not separate savings for each add-on; see how savings are measured.

Install Headroom for your operating system.

Not ready to install yet?

Leave your email and we'll send a one-page summary of the benchmarks, plus a note when big updates ship. No drip campaign.