@enconvert/mcp

Web and file reading for AI agents with a 0-1 render_quality score on every read. EnConvert perceives pages, discovers URLs, searches the live web, extracts structured data, ingests RAG-ready chunks and watches for changes, all through one API key, and every web read comes back scored with named deductions so a blocked, challenge or empty-SPA page is flagged instead of mistaken for content.

Official Model Context Protocol (MCP) server for EnConvert. It lets any MCP-compatible AI agent (Claude Code, Cursor, Windsurf, Claude Desktop, Continue, etc.) read the web and convert files through a single command.

Listed on: Official MCP Registry · Glama · LobeHub · npm

What you get

Twenty-four tools, optimized for reliable LLM tool selection. All URL/browser work goes through the V2 tools (perceive_url and friends); the V1 tools cover local/remote FILE conversion. All 24 are registered by default; see Toolsets to register only the web or only the file tools.

Every web read is scored

perceive_url (and everything built on it) returns render_quality (0.0-1.0) with named deductions such as anti_bot_challenge, http_error or login_wall, plus the source status_code. The contract your agent can rely on:

A detected content-free block is HTTP 200 with is_blocked: true, empty outputs and billed: false. The tool text says Blocked: yes (not billed). There is nothing to read; do not retry the same URL expecting a different answer.
Reads whose deductions include http_error or login_wall are also not billed (billed: false, text Not billed.). Cache hits are billed.
Treat render_quality below 0.40 as not-content: surface the score and deductions instead of the markdown.
A 429 comes back as plain text: Rate limited; retry after N seconds, from the gateway's Retry-After header.

File conversion (6)

Tool	What it does
`convert_document`	Convert DOC(X), XLS(X), PPT(X), ODT, ODS, ODP, OTS, Pages, Numbers, EPUB, HTML, MD, CSV, JSON, XML, YAML, TOML to PDF (or between each other)
`convert_image`	Convert between JPEG, PNG, SVG, HEIC, WebP, plus PDF → JPEG rasterization (handy for iPhone HEIC → WebP). SVG rasterization takes optional `width` and `height` controls
`compress_image`	Shrink a PNG, JPEG or WebP without changing its format, optionally to a target size in KB
`convert_anything_to_markdown`	Turn PDF, Office, ODF, EPUB, HTML, CSV or text files into clean Markdown for RAG pipelines
`convert_anything_to_pdf`	Turn almost any file (Office, ODF, iWork, images, SVG, HTML, Markdown, EPUB, RTF, CSV) into a PDF
`get_job_status`	Check a single file-conversion job by its job ID

Compression and sizing

compress_image never changes the format: a PNG comes back as a PNG, a JPEG as a JPEG, a WebP as a WebP, and the result is never larger than the input. It goes lossless first (metadata stripped, ICC profile and EXIF orientation preserved), then downscales with the aspect ratio locked if target_size_kb is still out of reach. An unreachable target_size_kb is not an error: you get back the smallest file achieved, so check file_size on the result. Animated APNG and animated WebP are rejected.

width and height on convert_image apply to SVG input only (svg to png, jpeg or webp; SVG to HEIC does not take them), 1 to 10000 each. Pass one alone and the other dimension is derived from the SVG's own aspect ratio. Pass both to set the exact canvas, which may change the aspect ratio. Omit both and the SVG's intrinsic size is kept.

V2: web data for agents (18, private API key required, plan-gated)

Tool	What it does
`perceive_url`	Render a page into multiple artifacts at once (markdown, HTML, screenshots, PDF, links, images) + structured extraction, with ~1h caching
`get_perceive_operation`	Re-fetch a perceive result with freshly signed artifact URLs
`perceive_batch`	Perceive up to 1000 URLs with shared options (inline for small batches, async for large)
`get_perceive_batch`	Poll a perceive batch by job ID
`discover_urls`	Enumerate a site's URLs via sitemap/crawl/hybrid, with no rendering and no render quota
`web_search`	Google-backed search (web, news, images, scholar, patents, maps) with optional auto-perceive of top results
`extract_structured`	Schema-driven data extraction (free CSS pass + LLM escalation) from up to 50 URLs or a discovered site
`start_ingest`	Turn a site or URL list into RAG-ready chunked JSONL (always async)
`list_ingest_jobs` / `get_ingest_job` / `cancel_ingest_job`	Manage ingest jobs
`retry_ingest_webhook`	Re-deliver a completed job's HMAC-signed completion webhook
`create_watcher` / `list_watchers` / `get_watcher` / `get_watcher_snapshots` / `update_watcher` / `delete_watcher`	Monitor pages for changes on an hourly-plus cadence with diff history and notifications

Not exposed by design: the webhook signing-secret endpoints (GET/POST /v2/ingest/webhook-secret*), because secrets do not belong in LLM context. Fetch those from the EnConvert dashboard.

Prompts (7)

Invokable workflows, surfaced as /mcp__enconvert__<name> in Claude Code. Each encodes ordering and cost rules the individual tools cannot enforce on their own.

Prompt	What it does
`site_to_rag_corpus`	Discover a site's URLs, filter them, ingest to chunked JSONL, poll to completion
`monitor_page`	Render once to pick the right diff mode, check for duplicates, then create the watcher
`convert_folder`	Route every file in a directory to the correct conversion tool by extension
`research_with_sources`	One `web_search(perceive_top=N)` round trip, then an answer with cited sources
`extract_dataset`	Write the schema, try the free CSS pass, then extract and report the LLM escalation
`diagnose_render`	Escalation ladder for an empty, partial or blocked render
`watch_digest`	Roll up every watcher's recent changes plus a health audit

Resources (4)

Read-only context addressed by URI. No resource read triggers a browser render, so attaching them costs no quota.

Resource	What it holds
`enconvert://formats`	The full conversion matrix: every implemented pair and accepted extension per tool. Offline, works without an API key
`enconvert://watchers`	Every page being monitored, with cadence, check counts and error streaks
`enconvert://ingest-jobs`	Recent ingest jobs with status, chunk counts and output links
`enconvert://watcher/{watcher_id}/snapshots`	One watcher's check history and diffs

Requirements

Node.js ^20.20.0 || >=22.22.0 (matches the package's engines field)
An EnConvert API key. Get one at enconvert.com/dashboard/api-keys

Install in one command

npx @enconvert/mcp setup

That's it. The wizard:

Detects your AI tools (Claude Code, Claude Desktop, Cursor, Windsurf, VS Code, Zed, Gemini CLI, Codex CLI, OpenCode) and lets you pick which get EnConvert. Detected tools are preselected.
Asks for your secret API key once (hidden input) and validates it live against the API. It even tells you if you pasted a public key by mistake.
Writes every config correctly, including the cmd /c npx wrapper Windows needs. Your key is stored once in ~/.enconvert/config.json (permissions 600), never copied into client configs.

$ npx @enconvert/mcp setup

  EnConvert MCP - setup

? Which AI tools should get EnConvert?
  [x] Claude Code (detected)    [x] Cursor (detected)
  [ ] Claude Desktop            [ ] Windsurf   ...
? Paste your SECRET API key (sk_live_..., input hidden): ********
  ✔ API key is valid.
  + Claude Code - claude CLI (user scope)
  + Cursor - ~/.cursor/mcp.json

  Done. Restart your AI tools to pick up the server.

Restart your AI tools afterwards, and you're running.

Manage it just as easily

Command	What it does
`npx @enconvert/mcp status`	Where it's installed + whether your key is valid (live check)
`npx @enconvert/mcp rotate-key`	Swap in a new API key: one command, applies to every client
`npx @enconvert/mcp upgrade`	Check npm for a newer version and upgrade (add `--dry-run` to preview)
`npx @enconvert/mcp remove`	Uninstall from selected tools (optionally delete the saved key)
`npx @enconvert/mcp setup --yes`	Non-interactive: configure all detected tools with the saved key

Scripting? setup --clients claude-code,cursor --api-key sk_live_... --yes skips all prompts. rotate-key with no argument prompts with hidden input so the key never lands in your shell history.

Where the key lives

setup stores your key once in ~/.enconvert/config.json (mode 600) instead of pasting it into every client's plaintext config. The server reads it at startup; the ENCONVERT_API_KEY environment variable always overrides it (for Docker, CI, or manual setups). Rotating is therefore a single-file change, and every client picks it up on its next launch.

Advanced: manual configuration

Prefer to wire it yourself? Add this to your client's MCP config (~/.cursor/mcp.json, ~/.codeium/windsurf/mcp_config.json, etc.):

{
  "mcpServers": {
    "enconvert": {
      "command": "npx",
      "args": ["-y", "@enconvert/mcp@latest"],
      "env": {
        "ENCONVERT_API_KEY": "sk_live_your_key"
      }
    }
  }
}

On native Windows, use "command": "cmd" and "args": ["/c", "npx", "-y", "@enconvert/mcp@latest"], because bare npx hangs. (WSL behaves like Linux.) For Claude Code:

claude mcp add enconvert -s user \
  -e ENCONVERT_API_KEY=sk_live_your_key \
  -- npx -y @enconvert/mcp@latest

The inline env block is optional if you've run setup (or created ~/.enconvert/config.json); the server falls back to the saved key automatically. Client config locations: Claude Desktop ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) / %APPDATA%\Claude\claude_desktop_config.json (Windows); Cursor ~/.cursor/mcp.json; Windsurf ~/.codeium/windsurf/mcp_config.json; Gemini CLI ~/.gemini/settings.json; Codex ~/.codex/config.toml.

Try it

Once installed, try any of these in a fresh prompt:

"Give me https://en.wikipedia.org/wiki/Model_Context_Protocol as markdown and summarize it." → perceive_url
"Screenshot https://news.ycombinator.com and save the page as a PDF too." → perceive_url (both artifacts in one call)
"Search for the three best static site generators and read their homepages." → web_search with perceive_top
"Get every plan name and price from https://example.com/pricing." → extract_structured
"Convert /Users/me/Desktop/report.docx to PDF." → convert_document
"Convert /Users/me/Desktop/iphone.heic to webp." → convert_image
"Squeeze /Users/me/Desktop/screenshot.png down to under 200 KB." → compress_image
"Turn /Users/me/Docs/whitepaper.pdf into Markdown for my RAG index." → convert_anything_to_markdown
"Watch https://example.com/changelog and tell me when it changes." → create_watcher

The agent picks the right tool automatically. Descriptions are tuned so URL work lands on perceive_url and file work on the convert tools.

Configuration

The API key is resolved in this order:

ENCONVERT_API_KEY environment variable (from your MCP client's env block), which always wins
~/.enconvert/config.json, written by npx @enconvert/mcp setup

Setting	Required	Default	Purpose
`ENCONVERT_API_KEY` (env) or `api_key` (config file)	✅ Yes	none	Your secret EnConvert API key
`ENCONVERT_BASE_URL` (env) or `base_url` (config file)	No	`https://api.enconvert.com`	Override for staging or self-hosted EnConvert
`ENCONVERT_TOOLSETS` (env)	No	`all`	Which tool family to register: `web`, `files` or `all`

Toolsets

Every registered tool costs context on each turn, so trim the set when you only need one family:

`ENCONVERT_TOOLSETS`	Tools	Use when
`web`	the 18 V2 tools: perceive, discover, search, extract, ingest, watch	the agent reads the web; nothing on disk to convert
`files`	the 6 V1 file tools: convert, compress, job status	a pure document-conversion assistant
`all` (default)	all 24	you want both

Set it in the client's env block next to the API key, e.g. "ENCONVERT_TOOLSETS": "web". The server logs [enconvert-mcp] toolsets=all: 24 tools registered to stderr on startup (stdout is the MCP transport). An unknown value is a startup error, not a silent default. Prompts and resources are always registered.

How it works

The server is a thin MCP wrapper around the official @enconvert/node-sdk, which owns all HTTP, auth, timeout (5 min), and job-polling fallback logic. Each tool handler maps MCP input to an SDK call and returns a consistent response:

a text summary with the download URL(s) and metadata
a structuredContent block with the full typed result
a resource_link to the local file when save_to is provided (file tools)
for perceive_url, the markdown artifact inlined in the response (up to ~256 KB) so agents can immediately read it without a second HTTP fetch

Troubleshooting

npx hangs on native Windows Use cmd /c npx …. The cmd shim handles Windows path resolution that bare npx does not. (npx @enconvert/mcp setup writes this automatically on Windows.)

Authentication failed: Invalid or missing API key Run npx @enconvert/mcp status. It shows where your key comes from and validates it live. Fix with npx @enconvert/mcp rotate-key, or check the ENCONVERT_API_KEY env block if you configured manually.

Tool call times out Browser renders of heavy pages (perceive_url, extract_structured) can take 15-30 s, and the SDK waits up to 5 min. If your client has a shorter timeout, raise it.

"Relative path" error on the file tools Pass an absolute path (e.g., /Users/me/file.docx or C:\Users\me\file.docx), or pass an http(s):// URL. MCP servers have no reliable working directory.

How it depends on the Node SDK

This MCP server is a thin wrapper around @enconvert/node-sdk, which owns all HTTP, authentication, timeout, retry, and job_id polling-on-500 logic. New SDK releases automatically benefit the MCP server.

License

MIT

EnConvert

Links

README