Content Core

What lfnovo/content-core MCP server does

The lfnovo/content-core MCP server gives MCP-compatible applications two content-processing tools: extract_content and summarize_content. Extraction accepts URLs, files, or text and covers web pages, YouTube videos, Reddit posts, PDFs, DOCX, PPTX, XLSX, EPUB, HTML, Markdown, plain text, audio, and video. The lfnovo/content-core MCP server returns plain text through the MCP interface.

Extraction is intended to hide differences between source types. URL handling can use automatic selection or a configured engine such as simple extraction, Firecrawl, Jina, or Crawl4AI. Document handling supports automatic or simple processing, with Docling available as an optional advanced parser. Audio and video processing can use speech-to-text providers through the project’s provider abstraction.

How it works

The package is implemented as a Python application with three access paths: an async Python API, the content-core command-line interface, and the MCP server. Starting content-core mcp launches the MCP process for a local client such as Claude Desktop.

The extraction tool can receive a URL, file path, or content value. The summarization tool works with text and can receive additional context that guides the requested summary. Configuration selects the URL and document engines, controls audio transcription concurrency, and identifies LLM or speech-to-text providers. Settings are resolved from constructor or command options first, then CCORE_* environment variables, the configuration file at ~/.content-core/config.toml, and defaults.

The lfnovo/content-core MCP server relies on external provider credentials when a selected engine or model requires them. Supported provider families include OpenAI, Anthropic, Google, Groq, DeepSeek, and Ollama, although the exact available options are provided by the Esperanto dependency.

Setup and configuration

Install the package with pip install content-core, or run it without a persistent installation through uvx. The MCP command is:

uvx content-core mcp

A Claude Desktop configuration can run that command and pass provider credentials such as OPENAI_API_KEY in the process environment. The project also supports persistent settings through content-core config set, config list, and config delete.

Relevant configuration names include CCORE_URL_ENGINE, CCORE_DOCUMENT_ENGINE, CCORE_AUDIO_CONCURRENCY, CCORE_LLM_PROVIDER, CCORE_LLM_MODEL, CCORE_STT_PROVIDER, CCORE_STT_MODEL, CCORE_STT_TIMEOUT, and CCORE_YOUTUBE_LANGUAGES. External services use their own standard variables, including OPENAI_API_KEY, FIRECRAWL_API_KEY, and JINA_API_KEY. Optional extras add Docling, Crawl4AI, or LangChain support. Crawl4AI may also require Playwright installation when used in local browser mode.

Tools and capabilities

• extract_content: Reads a URL, file, or text value and extracts usable content.
• summarize_content: Produces a summary from supplied text, with optional context.
• Automatic URL engine selection, with explicit engine overrides available.
• Document parsing for common office, ebook, markup, and plain-text formats.
• Audio and video transcription through configurable speech-to-text providers.
• Local MCP execution through uvx or an installed content-core command.

Limitations and notes

The MCP tools return plain text rather than a documented structured response schema. Some formats and engines depend on optional packages or external services. Selecting docling without installing the corresponding extra raises a configuration error; the default automatic document mode can fall back without it. Provider-based summarization, transcription, Firecrawl, and Jina use credentials and service availability outside the MCP process.

The lfnovo/content-core MCP server is distributed under the MIT license. The repository also includes an agent skill for Claude Code and Codex, but that skill is separate from the MCP transport and does not change the two MCP tools exposed by the server.