CVX v1.4.0 — plug it into your agent

Any agent — Claude, Cursor, VS Code, or any MCP client — now drives CVX end to end, locally. And any chat assistant can take a CV from one pasted prompt.

New

• MCP server — npx @hrtips/cvx mcp exposes four tools (get_schema, init_cv, validate_cv, build_pdf) wrapping the same engine as the CLI. No API keys, fully offline; the server teaches the model the edit→validate→build loop and the truthfulness rules.
• One-command client setup — cvx mcp init --client claude|claude-desktop|cursor|vscode writes/merges the right config shape per client (version-pinned, never clobbers other servers).
• The two-line front door — paste into any AI chat and you're done:

  Create my CV with https://github.com/hrtips/cvx — open that page and follow its instructions for AI assistants.
  Here is my LinkedIn profile: <your profile URL> (or attach your current CV and a photo instead).
  

  The docs carry the assistant the rest of the way: capability self-routing, the LinkedIn Save-to-PDF fallback, a non-technical handoff, and a review→brainstorm→preview step so nothing lands on your CV that you haven't seen.
• Agent Skill — skills/cvx/SKILL.md ships in the package for skill-capable agents.

Fixed

• Silent PDF corruption on forced pagination (found by dogfooding): a page1ExperienceCount that didn't fit compressed every text box into overlapping glyphs while the build reported success. Overflow now clips visibly at the page edge, and validate/build/build_pdf warn with calibrated estimates and concrete fixes. build --json and the MCP build_pdf tool now return a warnings array.
• Pagination semantics documented: entry N+1 starts page 2; automatic packing never overflows.

Quality gates

120 tests on ubuntu/macos/windows × Node 20/22/24; packaged E2E now includes an MCP initialize handshake against the packed tarball; reproducible builds stay byte-identical; install stays lean (tarball ~330 kB).