Open Source Radar — October 20: Headroom, HyperFrames, AI job search
Today's board is four tools that each delete a job from the agent loop: fewer tokens in, video out, applications drafted, diagrams drawn. Not one of them is a model — the open-source energy this week is all in the harness.
headroom (Python, ~76k stars) — A compression layer that sits between your agent and the model: tool outputs, logs, RAG chunks and files get squeezed before they reach the context window, while the originals stay cached locally so the agent can pull the detail back when it actually needs it. The project claims roughly 20% fewer tokens for coding agents and 60–95% fewer on JSON-heavy work, with the same answers. It ships as a library, a local proxy or an MCP server, and wraps Claude Code, Codex, Cursor and Aider without code changes. Reversibility is the design choice that matters: most compression tools throw information away and hope, this one lets the model ask for the original back. If your agent is burning its budget on stack traces and log dumps, start here.
HyperFrames (TypeScript, ~61k stars) — HeyGen's open-source framework that turns HTML, CSS and seekable animations into deterministic MP4s: the renderer seeks each frame in headless Chrome and encodes the result with FFmpeg, so identical input produces an identical video. That determinism is what makes video renderable inside CI and regression tests. It ships 21 skills agents load on demand — product-launch clips from a URL, pull-request explainers, faceless explainers, slide decks — so the authoring model is simply "an agent writes a web page, the renderer makes it move." It sits where Remotion sits, and its bet is the right one for this audience: plain HTML over React components, because the thing writing your composition is a model that already writes HTML well.
ai-job-search (Python, ~58k stars) — A Claude Code workflow that turns a job hunt into a pipeline: build a profile, scrape postings, score the fit, tailor the CV in LaTeX, draft the cover letter, prep for the interview. The author is a geophysicist whose position was cut in late 2025; he ran 69 tailored applications through it, got 20 first interviews and one signed contract, and started as an AI engineer in June 2026, per the project's README. The scrapers are built for the Danish market, so outside Denmark the reusable part is the structure, not the connectors — swap in your local job boards. It earns its keep at the 40 minutes per application, not at blasting 500 of them.
diagram-design (HTML, ~59k stars) — An agent skill that draws more than 40 diagram types — architecture, sequence, quadrant, timeline — as a single self-contained HTML and SVG file instead of another Mermaid grey box. It will pull your site's palette and fonts so every diagram after the first matches the brand, and the output opens by double-click with no build step and no rendering service. The anti-Mermaid stance is the entire pitch: Mermaid is fine for a README and indefensible in a deck.
Worth watching this week: whether token compression holds up when the detail matters. The failure mode is a model confidently answering from a summary that dropped the one line it needed.
Do you put a compression layer in front of your agent, or just eat the token bill? Write to us via our contact page.
