contentrx-mcp 0.7.0

ContentRX MCP server — content-design review for Claude Code, Cursor, and any MCP client.

contentrx-mcp

ContentRX as an MCP server — content-design review for Claude Code, Cursor, Claude desktop, and any other MCP client.

This is the surface that turns ContentRX from "a thing designers run on Figma frames" into "a thing your AI agent consults before writing a button label." It speaks Model Context Protocol over stdio and exposes seven tools — three for content review and four for Team-plan custom-example curation — all backed by the public ContentRX API.

Install

# One-line install + run via uv
uvx contentrx-mcp

# Or install in a project venv
pip install contentrx-mcp
contentrx-mcp

The server speaks MCP over stdio; you don't run it standalone — your MCP client launches it.

Configure your MCP client

You need a ContentRX API key (cx_...). Generate one at contentrx.io/dashboard — it's shown once at mint time, so save it before closing the page.

Claude desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or the equivalent on Linux/Windows:

{
  "mcpServers": {
    "contentrx": {
      "command": "uvx",
      "args": ["contentrx-mcp"],
      "env": {
        "CONTENTRX_API_KEY": "cx_your_key_here"
      }
    }
  }
}

Restart Claude desktop. The tools appear in the tools picker.

Claude Code

claude mcp add contentrx -- uvx contentrx-mcp
# Then set the env var in your shell or in the project's mcp config:
export CONTENTRX_API_KEY=cx_your_key_here

Cursor

Cursor's MCP config (Settings → MCP) takes the same shape as Claude desktop's. Use the same command / args / env block.

Tools

Content review

evaluate_copy — full review (counts against quota)

Check UI copy against the content-design standards library.

evaluate_copy(
  text: str,                      # the string to check
  moment_hint: str | None,        # optional: "error_recovery", "onboarding", etc.
  context: str | None,            # optional free-text context (reserved)
) -> {
  verdict: "pass" | "violation" | "review_recommended",
  review_reason: str | None,      # populated when verdict == "review_recommended"
  violations: [{
    issue: str,                   # human-readable description of the violation
    suggestion: str,              # concrete rewrite recommendation
    severity: "low" | "medium" | "high",
    confidence: float,            # 0..1
  }],
  warnings: [str],                # advisory; may be empty
}

Schema 2.0.0 envelope. The taxonomy is intentionally private — rule citations, internal IDs, and reasoning chains stay server-side. The public surface is the violation itself plus a concrete suggestion.

Counts against your monthly quota (Free: 25, Pro: 5,000, Team: 5,000 × seats).

evaluate_copy_batch — multiple strings in one call

Same review as evaluate_copy but takes a list. A dry_run flag lets the caller confirm quota cost before committing to a large batch.

evaluate_copy_batch(
  strings: list[str],
  moment_hint: str | None,         # applied to every string
  content_type_hint: str | None,   # applied to every string
  dry_run: bool = false,            # set true first for batches of 10+
) -> {
  # dry_run=true:
  dry_run: true, string_count: N, would_use_checks: N, message: str

  # dry_run=false:
  results: [
    { text: str, verdict, review_reason, violations, warnings } |
    { text: str, error: { kind, message } }
  ],
  checks_used: int,                 # successfully completed
  terminated_early: bool,
  termination_reason: str | None,   # set on auth/quota failures that abort the rest
}

Each successful string consumes one quota unit.

classify_moment — quick moment probe (no quota cost)

Classify what UI moment a string represents — without running the full evaluation. Useful for planning copy before you write it.

classify_moment(text: str) -> {
  content_type: str,              # e.g. "confirmation"
  moment: str,                    # e.g. "completing_task"
}

Free of quota. Rate-limited at 60/min per user (same bucket as evaluate_copy).

Team-plan custom examples

Team-plan teams can curate strings whose verdict is well-known to them — a phrasing they've already vetted as on-brand (verdict: "pass") or a known anti-pattern they don't want regressing (verdict: "violation"). When /api/check sees a matching string from that team again, it short-circuits to the stored verdict instead of re-running the LLM. Reduces noise on recurring voice calls and saves quota.

These tools are Team plan only and require team-admin role.

custom_example_add

custom_example_add(
  text: str,                       # the exact phrasing to short-circuit
  verdict: "pass" | "violation",
  moment: str | None,              # scope match to one moment context
  content_type: str | None,        # scope match to one content_type context
  standard_id: str | None,         # required when verdict == "violation"
  notes: str | None,               # 1–3 sentences explaining the decision
  contribute_upstream: bool = false,  # opt in to anonymised contribution
)

custom_example_list(limit?)

List the team's entries. Read-only; any authenticated team member can view.

custom_example_search(text)

Look up by normalised text. Use before _add to avoid duplicate-entry errors — matching is case + whitespace insensitive.

custom_example_remove(example_id)

Delete an entry by id.

Prompts

/review_ui_copy [focus?]

Multi-step review workflow. Walks every UI string in a file or diff through classify_moment → evaluate_copy, then summarizes violations by severity with suggested rewrites.

/review_ui_copy                              # uses file/diff in context
/review_ui_copy src/app/dashboard/page.tsx   # focus a specific file
/review_ui_copy "Click here to learn more"   # focus a single string

Appears as a slash command in Claude desktop and Cursor.

Examples

Three concrete scenarios showing how this surface fits into an AI- assisted coding workflow.

1. Catching a vague CTA before it ships

You ask Claude Code to add a button to a React component:

You: Add a "Click here to view pricing" button to the hero section.

Claude Code calls evaluate_copy before writing the JSX:

evaluate_copy(text="Click here to view pricing")

→ verdict: "violation"
  review_reason: null
  violations: [
    {
      issue: "Vague CTA — 'Click here' doesn't name the destination",
      suggestion: "Lead with the action verb + object: 'View pricing'",
      severity: "high",
      confidence: 0.91
    }
  ]
  warnings: []

Claude writes <Button>View pricing</Button> instead. Review that would have happened in a PR now happens before the first commit.

2. Reviewing a whole component file via /review_ui_copy

In Claude desktop with a file open in context:

/review_ui_copy src/app/dashboard/subscription-panel.tsx

The prompt walks every UI string in the file: classify_moment → evaluate_copy. Returns a structured summary grouped by severity with suggested rewrites. Typical output for a 200-line dashboard component surfaces 3–8 violations most human reviewers miss on skim.

3. Planning copy for a destructive-action dialog

You ask Claude for a confirmation dialog before shipping:

You: Write the copy for a "delete API key" confirmation.

Claude calls classify_moment first so the content-type signal is locked in before drafting:

classify_moment(text="Are you sure you want to delete this key?")
→ content_type: "confirmation"
  moment: "destructive_action"

Then evaluate_copy on candidate phrasings. The destructive_action moment weighs CLR-02 (consequence clarity) and TRS-01 (reversibility) more heavily — if the candidate doesn't mention "cannot be undone," ContentRX flags it and Claude revises before you see the draft.

Environment variables

Var	Default	Purpose
CONTENTRX_API_KEY	required	Your cx_... token from the dashboard
CONTENTRX_API_URL	https://contentrx.io	Override for local dev or self-hosting
CONTENTRX_INSECURE_HTTP	unset	Set to 1 to allow http:// for local dev — refuses otherwise so a typo can't leak the token

Errors

The server returns structured errors instead of stack traces. Every tool result is either a normal payload or:

{
  "error": "Rate limit hit. Try again in 30s.",
  "kind": "RateLimitError",
  "retry_after_seconds": 30
}

Error kinds

kind	What it means	Recommended action
AuthError	CONTENTRX_API_KEY is missing or malformed	Stop. Prompt the user to generate a key at https://contentrx.io/dashboard and set the env var.
AuthFailedError	The API rejected the cx_... token (revoked, rotated, or typo)	Stop. Prompt the user to re-mint their key at the dashboard.
QuotaExhaustedError	The user's monthly quota is at zero	Stop. Surface the included upgrade_url to the user. Don't retry.
RateLimitError	Per-user sliding-window rate limit hit	Wait retry_after_seconds, then retry once. If the second attempt also 429s, surface the error and let the user decide.
ContentRXError	Generic upstream failure (5xx, network blip, unexpected response shape)	Retry once with a short backoff (1–3s). If still failing, surface the error.

MCP clients can branch on kind to retry, prompt the user to upgrade, or stop. The error string is human-readable but not machine-stable — always key off kind for control flow.

Development

cd mcp-server
pip install -e ".[dev]"
pytest tests/

# Local dev against npm run dev:
export CONTENTRX_API_URL=http://localhost:3000
export CONTENTRX_INSECURE_HTTP=1
export CONTENTRX_API_KEY=cx_...     # mint via npm run dev dashboard
contentrx-mcp                        # speaks MCP over stdio

The MCP CLI from mcp[cli] is useful for poking the server without a real client:

mcp dev contentrx-mcp                # opens the inspector UI

Release checklist

See mcp-server/CLAUDE.md for the version-bump + PyPI publish flow.

License

MIT — same as the ContentRX project.