Setup — one-time per machine

1. API keys via ~/.secrets

This skill never sees a key in source. Keys live in ~/.secrets (gitignored, not in any dotfiles repo) and are inherited by Claude Code from the shell at startup.

~/.secrets (chmod 600):

# Crawl, fetch, ingest skill
export FIRECRAWL_API_KEY="fc_..."
export TAVILY_API_KEY="tvly-..."
export OPENGRAPH_IO_API_KEY="..."
export JINA_API_KEY="jina_..."     # optional, raises Jina Reader rate limits

# Brand-asset cascade (logo hunt)
export BRANDFETCH_API_KEY="..."    # optional but recommended; free tier 1k req/mo
export GOOGLE_CSE_KEY="..."        # optional, last-resort SVG search via fileType=svg
export GOOGLE_CSE_CX="..."         # optional, CSE engine ID configured for image search

~/.zshenv adds one line:

[ -f ~/.secrets ] && source ~/.secrets

After editing, restart Claude Code so MCP server subprocesses inherit the new env vars. Existing sessions started before the keys were added will not see them.

2. MCP servers

Two MCP servers are added at user scope (apply to every project):

claude mcp add -s user firecrawl -- npx -y firecrawl-mcp
claude mcp add -s user tavily -- npx -y tavily-mcp

Both servers read their API keys from process.env (inherited from your shell), so no -e KEY=value flag is needed when ~/.secrets is sourced.

Verify:

claude mcp list
# Should show: firecrawl ✓ Connected, tavily ✓ Connected

If a server fails to connect with an auth error, the env var isn’t reaching it — check that ~/.zshenv sources ~/.secrets and that you’ve restarted Claude Code.

3. Helper scripts

scripts/og-fetch.ts (OpenGraph.io) and scripts/jina-reader.ts (Jina Reader) are run via pnpx tsx. The shebang handles this automatically:

chmod +x scripts/og-fetch.ts scripts/jina-reader.ts
./scripts/og-fetch.ts https://stripe.com
./scripts/jina-reader.ts https://www.sequoiacap.com/our-companies/

tsx is fetched on demand by pnpx; no global install required.

4. Local CLIs for asset processing

ImageMagick (required for scripts/bg-strip.sh tier 1)

brew install imagemagick
magick -version | head -1   # should print: Version: ImageMagick 7.x ...

Tier 1 of the background-strip cascade is pure ImageMagick — flood-fill from each corner with the sampled corner color (handles white and arbitrary brand-color backgrounds). Deterministic, fast, no model deps.

rembg (optional, for scripts/bg-strip.sh tier 2)

brew install pipx           # if not already installed
pipx install rembg[cli]
rembg --help                # smoke test; first real call downloads ~170MB model

Tier 2 of the background-strip cascade — U²-Net segmentation model, runs locally, MIT-licensed. Handles non-uniform backgrounds, drop shadows, soft edges. Skip this install if disk space matters; the bg-strip script gracefully degrades when rembg is missing (logs the case as rembg-unavailable in the company’s notes).

5. Cache directory

The skill writes raw API responses to:

~/.claude/skills/crawl-fetch-ingest/cache/{firm-slug}/
  firecrawl/<hash>.json
  tavily/<hash>.json
  jina/<hash>.md
  og/<hash>.json

Auto-created on first use. Hash is sha1(input).slice(0, 16) where input is the URL (Firecrawl, Jina, OG) or the query string (Tavily). To force re-fetch for a firm: rm -rf ~/.claude/skills/crawl-fetch-ingest/cache/{firm-slug}/.

6. Symlink (already done at scaffold time)

ln -s /Users/mpstaton/code/lossless-monorepo/context-v/skills/crawl-fetch-ingest \
      ~/.claude/skills/crawl-fetch-ingest

If the symlink is missing, recreate it. The skill source of truth is the context-v/skills/ repo; ~/.claude/skills/ is just where Claude looks.

7. New machine checklist

1. Clone context-v repo (with submodules)
2. Create ~/.secrets with the env vars (chmod 600)
3. Add [ -f ~/.secrets ] && source ~/.secrets to ~/.zshenv
4. Run the two claude mcp add commands above
5. brew install imagemagick
6. (Optional but recommended) brew install pipx && pipx install rembg[cli]
7. Symlink the skill into ~/.claude/skills/
8. Restart Claude Code
9. Verify with claude mcp list and magick -version