CV Builder

A self-hosted tool for maintaining a CV as structured content instead of a single hand-formatted document: a browser editor for a data-driven HTML/CSS resume, a SQLite-backed document and snippet library for assembling role-tailored variants, and an MCP server so an LLM can help do the same thing from a chat client instead of the browser.

Ships with a fully synthetic example person (Homer Simpson) in cv/web/data.yaml and content/ — replace it with your own before you rely on this for real. Nothing in this repo is anyone's real personal data.

Layout

• cv/web/ — HTML/CSS CV source (data.yaml, template.html.j2, style.css, editor.js) and the app UI (cv/web/src/: shared shell

  • design tokens in shell/ and theme.css, one Jinja page + CSS/JS per app-chrome route under pages/)

• content/ — optional additional detail for the snippet library, as markdown alongside the bootstrap YAML:

  • work-experience/ — one file per employer (category: experience, company taken from the filename)

  • parts/ — reusable blocks not tied to an employer, e.g. an alternate bio or a longer "detailed" variant of a strength (category: part)

  • requirements/ — pre-written answers to recurring posting requirements, for match_job_posting/the builder's posting-matcher to surface (category: requirement)

  • See the example files under each for the heading convention: only the last heading before a block of prose becomes a snippet

• src/cvbuilder/ — SQLite-backed CV document store, snippet library, importer, matcher, composer, exporters, and the MCP server

• scripts/ — CLI entry points (see below)

• data/ — local (non-Docker) SQLite database (snippets.db, gitignored); when CV_DATA_ROOT is set (Docker defaults to /data), the DB, uploads, imports, and export artefacts live under that data root instead

• Explicit variant export files default to cv/variants/ locally, or $CV_DATA_ROOT/cv/variants/ in Docker

Related MCP server: decroche-mcp

Quickstart

pip install -r requirements.txt

# Seed the SQLite database from cv/web/data.yaml + content/
PYTHONPATH=src python3 scripts/seed-snippets.py

# Run the editor / builder / variants UI
python3 scripts/serve-editor.py        # http://127.0.0.1:5057/edit

Requires google-chrome or chromium on PATH for PDF export (CHROME_BIN env var to point at a specific binary).

Document storage and export

Master and variant CV documents live in SQLite (cv_documents), which is the source of truth for live browser edits, composed variants, imports, and exports. On first run, if the database has no master CV row, the app bootstraps that row from the shipped cv/web/data.yaml; after that, data.yaml is just an input or explicit export target, not the live store.

Undo and redo state is transitory and stored in cv_history. Pins preserve frozen document content together with the snippet stacks used to produce it, so a pinned variant can be inspected or exported later without depending on current snippet selections.

Exports are explicit: request YAML, Markdown, or PDF when you want files on disk. Composing or editing a CV updates the database first and writes cv/variants/<name>/ output only when export options are selected.

Editing in the browser

scripts/serve-editor.py serves one app, sharing a common nav/header shell (cv/web/src/shell/) across every page below, including Working Draft (/edit):

• / — Home dashboard: live snippet/version counts and recent versions.

• /edit ("Working Draft") — click any text to edit it in place inside the shell; hover controls add/reorder/delete list items (bullets, skills, jobs, subsections, education, custom side panels); Save & Preview stores the Working Draft in SQLite and can render a real PDF. Adding a skill/bio paragraph/education entry opens a picker fed from the snippet database, with search and duplicate flagging.

• /build ("Tailor") — paste a job posting to re-rank snippets by keyword match, choose content, assemble an ordered draft, and compose it into a named SQLite variant; choose export options when you want YAML, Markdown, or PDF files under cv/variants/<name>/.

• /library ("Content library") — browse/search/filter every snippet, switch between its brief/standard/detailed variants, and create/edit/delete snippets or re-seed the database from source files.

• /variants ("Versions") — preview, re-render, export, or delete composed variant documents.

• /assets — browse/upload photos and logos (backed by the /api/images* endpoints) and reference the built-in contact icons.

• /connect ("Connect AI") — MCP setup instructions and an optional local connectivity check.

Docker

Runs the editor/builder UI and the MCP server in one container. Mutable user content is stored on a named volume at /data (cv_data), so the database, uploaded images, resume imports, and optional variant/PDF exports survive image rebuilds. The repo is still bind-mounted at /app for local code edits — omit that mount for image-only runs.

docker compose up --build
# home:     http://127.0.0.1:5057/
# editor:   http://127.0.0.1:5057/edit
# tailor:   http://127.0.0.1:5057/build
# library:  http://127.0.0.1:5057/library
# variants: http://127.0.0.1:5057/variants
# assets:   http://127.0.0.1:5057/assets
# connect:  http://127.0.0.1:5057/connect
# MCP:      http://127.0.0.1:8765/mcp  (streamable-http)

Volume layout (CV_DATA_ROOT=/data):

/data/snippets.db
/data/assets/images/     # uploaded photos and logos
/data/imports/           # resume uploads
/data/cv/variants/       # optional YAML/PDF exports
/data/cv/current/        # disposable preview artefacts

To use a host directory instead of the named volume, replace the cv_data:/data mount with e.g. ./persistent-data:/data.

Both ports publish to 127.0.0.1 only. Set ENABLE_MCP=0 in the compose environment block to run the web UI without the MCP server.

Connecting an LLM (MCP server)

src/cvbuilder/mcp_server.py (run via scripts/mcp-server.py) exposes the snippet library and composer as MCP tools: list_snippets, get_snippet, create_snippet, update_snippet, add_snippet_variant, delete_snippet, delete_snippet_variant, audit_library, upsert_snippets (batch create/update; dry_run defaults to true), delete_snippets (batch delete; dry_run defaults to true), match_job_posting, compose_cv, list_variants, list_drafts, get_draft, save_draft, delete_draft, reseed_snippets. Use audit_library then dry-run upsert_snippets / delete_snippets to populate or refine the Content library.

Local subprocess (stdio) — the client spawns the server itself:

claude mcp add cv-builder -- python3 /absolute/path/to/cv-builder/scripts/mcp-server.py

{
  "mcpServers": {
    "cv-builder": {
      "command": "python3",
      "args": ["/absolute/path/to/cv-builder/scripts/mcp-server.py"]
    }
  }
}

Already-running server (HTTP) — point a client at the Docker container instead (MCP_TRANSPORT=streamable-http inside the container, published at http://127.0.0.1:8765/mcp):

claude mcp add --transport http cv-builder http://127.0.0.1:8765/mcp

{
  "mcpServers": {
    "cv-builder": { "url": "http://127.0.0.1:8765/mcp" }
  }
}

There's no authentication on the MCP endpoint or the web UI — fine for local personal use, but don't publish either port beyond 127.0.0.1 without adding auth first.

Set SNIPPETS_DB to point either server at a different database file (defaults to data/snippets.db).

API endpoints (same Flask process as the editor)

• GET /api/person — read-only person block of the master CV (Assets page)

• GET/POST/PUT/DELETE /api/snippets — list/create/update/delete snippets

• DELETE /api/snippets/<id>/variants/<level> — remove one detail level

• POST /api/structure — insert/delete/move/replace items in the master CV

• GET /api/history, POST /api/undo, POST /api/redo — editor undo/redo

• GET/PUT/DELETE /api/drafts[/<name>] — saved builder drafts

• POST /api/match — rank snippets against posting text

• POST /api/compose — compose a named variant from selected snippet ids, with optional YAML, Markdown, and PDF exports

• GET/DELETE /api/variants[/<name>], POST /api/variants/<name>/render

• GET /api/images, POST /api/images/upload, POST /api/images/fetch — list, upload, or download images/icons into assets/images/

• POST /api/seed — re-seed the database from YAML + markdown sources

• POST /api/imports/<token>/confirm — confirm staged import (mode: library|master); master imports update SQLite and can be exported later

Tests

python3 -m pytest

Cucumber / Behave BDD

Gherkin features under features/ describe product behaviour informed by the interactive wireframe (cv/web/wireframe.html). Behave always runs against the shipped Flask app (test client + isolated SQLite / import scratch dirs) — never against the wireframe.

Tag @wip marks wireframe-informed backlog that is not yet asserted against the real UI (usually browser-driven flows). Those scenarios are skipped with a reason so the default suite stays green while gaps stay visible.

pip install -r requirements.txt
behave
# include skipped @wip backlog in the report (already shown by default):
behave --tags=@wip