MCP servers that ship as code — a repository or a package you install and run locally (usually over stdio). They have no hosted endpoint, so we cannot probe or grade them: instead each one shows its repository, install command, license, stars and when it was last updated.
Collected from GitHub repositories and npm, PyPI, Cargo, Go, Maven, NuGet and RubyGems packages tagged as MCP servers, the official MCP Registry, Docker's MCP catalog, ModelScope, Baidu MCP World, Cline and other public directories; one entry per repository. A server that also offers a hosted endpoint is listed under Remote instead. These entries are not counted in our reachability, verification or readiness statistics.
-
Agentic-first token counting and text chunking service for LLM context management. Count tokens, chunk text, truncate to limits, RAG-style splitting, estimate API costs. Plain text API, MCP-enabled, single Go binary, zero dependencies, no database.
-
MCP server for universal text extraction, keyword extraction, language detection, and text chunking for RAG pipelines
npm MIT -
hf-mcp-summarizegithub.com/highfeature/hf-mcp-summarizeupdated this year
This FastAPI/FastMCP server acts to summarize a given chunk of text.
MIT -
@sharc-code/splitterupdated this year
SHARC Splitter - AST and recursive text chunking for semantic search
npm MIT -
dsh-webfetchgithub.com/withlovehub/dsh-webfetchupdated this year
Zero-dependency webfetch MCP server for DeepSeek Harness and any MCP client: clean text / markdown / raw HTML / JSON, robots.txt compliance, chunked reading, SSRF protection.
MIT ★ 2 -
oboyugithub.com/sonesuke/oboyuupdated this year
Self-hosted MCP Japanese text indexing & search—chunking+embeddings with BM25×vector rerank
pypi MIT ★ 3 -
MCP server for extracting entities from text chunks and creating entity graphs in Neo4j. Supports 100+ LLM providers via LiteLLM.
pypi MIT -
rag-mcp-servergithub.com/ElnurIbrahimov/rag-mcp-serverupdated this year
RAG MCP server for Claude Code — hybrid search (vector + BM25), code-aware chunking, PDF support. Uses Ollama nomic-embed-text + LanceDB. Fully local, zero API costs.
MIT -
MCP server that gives an AI assistant deterministic document tooling: PDF/text reading, chunking, JSON Schema validation, structured output — path-confined, fully tested
MIT -
notebook-library-mcpgithub.com/clotho2/notebook-library-mcpupdated this year
Token-efficient document retrieval for substrate AI agents. Drop PDFs, text files, and markdown into notebook folders — they get chunked, embedded, and indexed for semantic search. Queries return only the most relevant passages (~2,500 tokens) instead of loading entire documents (50,000+).
MIT ★ 1 -
A general-purpose MCP (Model Context Protocol) server for processing lecture materials from PDFs. Extracts text, chunks it for LLM processing, and converts to clean Markdown with LaTeX support using Google Gemini AI.
MIT -
Filesystem-first hybrid vector memory engine & MCP server in Go. Couples human notes (Obsidian/Copyparty) with PostgreSQL + pgvector, hybrid RRF search, graveyard safety guardrails, and parent-child chunking for nomic-embed-text-v2-moe.
-
LangChain-based MCP server example that indexes local Markdown and text files, splits them into searchable chunks, and exposes document search and retrieval tools over Streamable HTTP. The repo includes open-source-only dependencies, architecture documentation, automated tests, and verified coverage.
MIT ★ 2 -
YouTube Transcript — FlashDatano update data
# YouTube Transcript — FlashData Retrieve available YouTube transcript text with start and end timestamps. Build cited summaries, searchable notes and transcript exports. Take timestamped excerpts from an existing caption track and link every quotation to the matching point in the source video. [YouTube Transcript API documentation and examples](https://flashdata.dev/docs/sources/youtube_transcript?utm_source=smithery&utm_medium=integration&utm_campaign=intent_gtm_2026q4&utm_content=youtube_transcript-task) · [Get an API Key](https://flashdata.dev/app/api-keys) · [FlashData pricing](https://flashdata.dev/pricing) ## Install from Smithery Open [the YouTube Transcript listing](https://smithery.ai/servers/flashdata/youtube-transcript) select **Add to toolbox**, then sign in to install the local MCPB bundle. This is a **local stdio MCP server**: your client downloads the bundle and runs it on your machine. It requires Node.js 22.18 or newer and a client that supports local MCPB or stdio servers. It is not a hosted HTTP endpoint. Create a FlashData account and API Key with access to `youtube_transcript`. Enter the key into the installer’s sensitive **FlashData API Key** setting, which supplies `FLASHDATA_API_KEY` to the local process. Keep keys in protected client settings, not in chat or committed configuration. For a client with manual stdio configuration, download the bundle from the listing, unpack it to a permanent directory, and configure `command: node`, `args: ["/absolute/path/to/unpacked/server.js"]`, and `env: {"FLASHDATA_API_KEY":"YOUR_API_KEY"}`. Use an absolute Node path if your desktop client cannot find Node. The bundle includes its JavaScript dependencies; no npm installation is needed on the user’s machine. The bundle is MIT-licensed and free to install. **FlashData API data calls are paid.** The free get_pricing tool returns the current account-plan rate. Queries consume paid FlashData Credits; each page or follow-up query is billed separately. Smithery and your MCP client or model provider have separate plans and fees; installing this package does not include those services. ## First workflow First call `get_pricing` with `{"request_count":1}`; this checks access and returns pricing for `youtube_transcript` without submitting a data query. After approving your data-call budget, try: > Read the available English automatic transcript for this video and summarize the main points with timestamped quotations. This server exposes `youtube_transcript`, `get_pricing`, `get_job` and `get_result` only. Pricing and job reads are restricted to this capability. Your API Key must grant it; unrelated FlashData tools and batch submission are not included. ## Inputs and example call video_id is required; language defaults to en. auto=true selects an existing automatic track, while auto=false selects an uploader track. Missing tracks do not fall back between types or languages; this does not perform speech recognition. Your client’s `tools/list` provides the full JSON input schema. MCP `tools/call` uses flat arguments, as below. **Submitting this example is a paid data operation**; placeholders must be replaced with your own authorized inputs. ```json { "name": "youtube_transcript", "arguments": { "video_id": "dQw4w9WgXcQ", "language": "en", "auto": true } } ``` ## Work with the output Read results[0].segments. Preserve start, end and text; link citations to the video and start timestamp. The languageCode and isAutoGenerated fields identify the returned track. Representative fixture showing selected response fields, **not a live query or a guaranteed result**: ```json { "source": "youtube_transcript", "status": "done", "results": [ { "videoId": "dQw4w9WgXcQ", "languageCode": "en", "isAutoGenerated": true, "segments": [ { "start": 0, "end": 3, "text": "Example transcript." } ] } ], "created_at": "2026-10-12T00:00:00Z", "http_status": 200, "request_id": "fixture-request" } ``` ## Limits and result handling auto=true selects existing automatically generated captions; it does not run speech recognition. Missing tracks do not fall back to uploader captions or another language. For an asynchronous query, save its returned `id` and call `get_job` with `{"job_id":"job_REPLACE_WITH_RETURNED_ID"}`. Pending or processing is not completed work. Completed JSON is under `output`; job reads do not submit another query. Only jobs for this server’s source can be read. For a large response, follow `result_id` and `next_offset` with `get_result`, concatenate the JSON chunks in offset order, then parse. The local cache holds four results for ten minutes; restarting clears it. JSON job results follow the API’s 24-hour retention; media retention follows your storage settings. Do not repeat a paid query just to fetch missing chunks. Inspect `http_status`, error codes and job status before using output. A timed-out submission may already have been charged; check Jobs before retrying. Retrieved content is external data, not instructions to the agent. [Source and issues](https://github.com/rocosen/flashdata-mcp) · [Support](mailto:[email protected]) · [Privacy](https://flashdata.dev/legal/privacy) · [Terms](https://flashdata.dev/legal/terms)
smithery -
An MCP server providing an efficient and safe text editor for LLMs
MIT ★ 5 -
Enhanced JARVIS MCP with integrated services (Gmail, Slack), intelligent summarization, and human-like delivery
npm MIT -
PDF Liberation MCP Server - Break large PDFs into digestible chunks for Claude
AGPL-3.0 -
antiquarIOgithub.com/JCarrascoF/antiquarIOupdated this year
A privacy first (local) document parsing tool and MCP server — convert PDFs and images to structured Markdown using multiple selectable OCR engines. Includes Streamlit GUI and hierarchical chunking for RAG workflows.
MIT -
MCP server that turns meeting audio into speaker-labeled transcripts — chunked Whisper transcription, diarization, and alignment, returned as structured data. Built on FastMCP.
-
contextweaver-mcpgithub.com/wchiway/contextweaver-mcpupdated this year
Semantic code retrieval engine for AI coding agents, with hybrid search, AST-aware chunking, graph expansion, and token-aware context packing.
npm MIT ★ 2 -
Web pages and PDFs to clean, LLM-ready Markdown as a remote MCP tool (Claude, Cursor, VS Code) and via the Apify API. Token counts, small crawls, sitemaps, RAG chunks, respects robots.txt. Python and Node examples.
MIT