cv-tailor-mcp

An MCP server that tailors a one-page LaTeX CV to a specific job posting without burning tokens on repetitive work: re-reading your whole CV source, writing the same LaTeX boilerplate over and over, or pasting pdflatex compiler logs into the chat.

It works identically from Claude Code and Cursor (or any other MCP-compliant client) -- MCP is an open protocol, so the same server binary is just registered in each client's own config file.

This repo contains no personal data. You keep your own cv_facts.yaml (your real experience/projects/skills) in your own private repo, and point this server at it via environment variables. See examples/cv_facts.example.yaml for a fictional but complete example, and docs/SCHEMA.md for the full schema.

What it does

Instead of:

1. Reading your entire CV source file into the model's context every time,
2. reading a previous tailored variant as a style reference,
3. having the model write ~150-250 lines of near-identical LaTeX by hand, and
4. pasting compiler output into the chat 3-5 times while manually tightening spacing to fit one page...

...you get 5 small tool calls:

Tool	Purpose
list_tags	Discover the tags in your cv_facts.yaml
get_facts(tags=[...])	Get only the relevant slice of your experience/projects/skills
list_variants	See CVs you've already generated (tagline, tags) without re-reading each file
render_cv(payload)	Generate cv_<variant>.tex from bullet/project/skill-group ids -- never fabricates, errors on unknown ids
compile_cv(variant)	Run pdflatex, get back {success, pages, first_error}, auto-cleans .aux/.log/etc.

There's no automatic "shrink until it fits one page" tool in v1 -- if compile_cv reports 2 pages, call render_cv again with spacing_profile: "compact" or "tight" (see docs/SCHEMA.md).

Install

Requires Python 3.10+ and a LaTeX distribution (pdflatex on your PATH, or set PDFLATEX_PATH explicitly).

git clone https://github.com/<you>/cv-tailor-mcp.git
cd cv-tailor-mcp
./setup.sh

This creates a local .venv/ and installs the server into it.

Set up your own CV data

1. Copy the example somewhere outside this repo (e.g. into your own private CV repo) and fill it in with your real information:

   cp examples/cv_facts.example.yaml ~/my-cv/cv_facts.yaml
   

2. Edit it -- see docs/SCHEMA.md for the full field reference. Only put things that are actually true; the point of render_cv's id-based lookup is that it can't fabricate content that isn't in this file.

Register the server

Both Claude Code and Cursor read the same shape of config -- just a different file path:

Claude Code -- .mcp.json in your project root (or ~/.claude.json for a user-wide server):

{
  "mcpServers": {
    "cv-tailor": {
      "command": "/absolute/path/to/cv-tailor-mcp/.venv/bin/python",
      "args": ["/absolute/path/to/cv-tailor-mcp/server.py"],
      "env": {
        "CV_FACTS_PATH": "/absolute/path/to/your/cv_facts.yaml",
        "CV_OUTPUT_DIR": "/absolute/path/to/where/you/want/generated/cvs"
      }
    }
  }
}

Cursor -- the same mcpServers block in .cursor/mcp.json (project) or ~/.cursor/mcp.json (global).

Environment variables

Variable	Required	Default
CV_FACTS_PATH	yes	--
CV_OUTPUT_DIR	yes	--
CV_TEMPLATE_PATH	no	bundled templates/default_cv_template.tex.j2
PDFLATEX_PATH	no	resolved via PATH (shutil.which("pdflatex"))

The server fails fast with a clear message on startup if a required variable is missing or points at something that doesn't exist -- it never silently falls back to guessing.

Example session

> list_tags
{"tags": [{"name": "backend", "experience": 1, "projects": 1, ...}, ...]}

> get_facts(tags=["backend", "cloud"])
{... only the entries/bullets tagged backend or cloud ...}

> render_cv({
    "variant": "acme",
    "tagline": "Backend Engineering Intern Candidate",
    "summary": "...",
    "skill_group_ids": ["languages", "backend_cloud"],
    "experience": [{"id": "acme_backend_intern", "bullet_ids": ["acme_api_bullet", "acme_pipeline_bullet"]}],
    "projects": [{"id": "proj_taskflow", "bullet_ids": ["taskflow_realtime_bullet"]}]
  })
{"tex_path": ".../cv_acme.tex", "line_count": 118, "spacing_profile": "default"}

> compile_cv("acme")
{"success": true, "pages": 1, "first_error": null}

Bring your own template

If the bundled one-page style doesn't match yours, write your own .tex.j2 (Jinja2, using << >> for variables and <% %> for blocks/loops instead of the default {{ }}/{% %}, since LaTeX already uses {/}) and point CV_TEMPLATE_PATH at it. It must accept the same context documented in docs/SCHEMA.md -- see templates/default_cv_template.tex.j2 for a working reference.

Scope / non-goals for v1

• No auto-fit-to-one-page spacing tool -- three manual spacing_profile presets instead (see above).
• cv_facts.yaml is a second source of truth: if your real experience changes, update it yourself. Nothing syncs it from anywhere automatically.
• compile_cv only confirms the page count pdflatex reports -- it doesn't check whether the result looks good. Open the PDF yourself before you send it anywhere.

License

MIT, see LICENSE.