Cross-source recommendation & sentiment for anything — "best running shoes", "how good is the new Claude model", "best ramen near Union Square". Vox searches Reddit + X + the web, cross-checks them, and returns a ranked, cited, honestly-hedged answer. Read-only; runs on your Claude subscription (no API).
./install.sh # symlinks skills/* into ~/.claude/skills/The web tier (WebSearch/WebFetch) is built in. The other tiers are capability-gated —
each needs its CLI/tool on PATH: Reddit→reddit-cli, X→bird, Maps→maps-cli(+gosom),
Video→tiktok-cli+mw+ffmpeg, browser/logistics→the claude-in-chrome MCP. Vox runs with
whatever is present and degrades honestly.
Setting up on a fresh machine? This is a multi-repo system (sibling CLIs + external tools). Follow SETUP.md end-to-end — it covers cloning the public repos, the sibling installs, external tools, MCP setup, credential re-auth, and verification.
In Claude Code: /vox best running shoes for flat feet under $150
A vox orchestrator skill probes sources, proposes a rubric, fans out one subagent per relevant
source — vox-reddit, vox-x, vox-web, plus the vox-maps (places), vox-browser
(logistics/place-data fallback), and vox-video (TikTok collections) tiers when the query needs
them — corroborates across sources (2+ to promote), resolves or discloses conflicts, and renders
one of two output skeletons (ranked finalists, or a no-finalist sentiment read) with per-figure
confidence. See docs/superpowers/specs/2026-06-07-vox-design.md and the mined docs/vox-recipe.md.
.venv/bin/python -m pytest
.venv/bin/python -m ruff check tools tests eval
.venv/bin/python tools/validate_skills.py skillsShipped since v1: the browser tier (vox-browser, Maps/logistics via Claude in Chrome), the
gosom-backed places tier (vox-maps), the TikTok collection analyzer (vox-video, with an
optional agy/Gemini cross-check), and Firecrawl as an anti-bot rung in vox-web. Deferred:
Instagram video ingest, a vox "..." shell wrapper, and private/own-collection TikTok access.