Contextator
Your docs, at one address your AI tools can read.
Brand: Tunedness
Contextator is a self-hosted MCP server that turns your documentation into one semantically searchable endpoint. Every project has its own MCP URL, and your AI coding tools reach the right documents through it. It is built by Tunedness.
One endpoint per project
Each project gets its own address in the form /mcp/<project> and is isolated end to end from the others. Cross-project search is absent on purpose. The server exposes the search_docs, list_topics, read_document, whats_new and report_gap tools; report_gap writes only to its own table. A new project requires a bearer token by default; open access or account-based access (OAuth 2.1) can be chosen instead.
Sources and search
Documents can come from a local directory, a git repository, archives, Notion, Confluence, documentation sites, S3-compatible buckets, Google Drive and GitHub issues. Search combines vectors with keywords; when the score stays below the threshold, Contextator says there is no good match instead of inventing a result. Indexing is incremental. OpenAPI files are indexed as one document per operation.
On your own server
The embedding model is Xenova/multilingual-e5-small; it runs on the CPU and covers one hundred languages, Turkish included. Once the model is cached, local sources cause no outbound traffic. It installs with Docker or Docker Compose: PostgreSQL and pgvector live inside the same container, there is a separate image for an external PostgreSQL and a Helm chart for Kubernetes. The dashboard opens at http://localhost:3444/ and generates connection settings for Claude Code, Cursor and Claude Desktop.
Limits
The project site lists its own limits: no OCR, weak cross-language retrieval, no cross-project search, and prompt injection is not filtered.
License
AGPL-3.0-or-later or a separate commercial license. Because the copyright is held by a single party, a commercial license can be granted where the AGPL does not fit. The source is developed on GitHub.
Key features
- A separate MCP URL for every project, isolated end to end
- Sources such as git repositories, Notion, Confluence, documentation sites, S3-compatible buckets, Google Drive and local folders
- Hybrid search that combines vectors and keywords; says so when there is no good match
- Works with Claude Code, Cursor, Claude Desktop and any MCP client
- Local by default: the embedding model runs on your own server
- One Docker container; a Helm chart for Kubernetes
Frequently asked questions
What is Contextator for?+
It indexes your documentation and serves it as a single MCP endpoint per project. Claude Code, Cursor or any other MCP client searches your docs semantically through that URL.
Which sources can it read?+
A local directory, a git repository (a token and push webhook for private repositories), uploaded archives, Notion, Confluence, documentation sites, S3-compatible buckets, Google Drive and GitHub issues. Markdown, MDX, plain text, HTML, Word, CSV and PDF files are read.
Does my data leave my servers?+
Not by default. The embedding model runs locally, and once it is cached, local sources cause no outbound traffic. You can switch to OpenAI embeddings with two environment variables. Source tokens are stored encrypted with AES-256-GCM.
Does it work on Turkish documentation?+
The embedding model covers one hundred languages, Turkish included. The project site states openly that cross-language retrieval is weak.
What are its known limits?+
There is no OCR for scanned documents, there is deliberately no cross-project search, and prompt injection inside documents is not filtered.
What is the license?+
AGPL-3.0-or-later or a separate commercial license. A commercial license can be granted where the AGPL does not fit; its price is not stated in public sources.
Let’s build your next project together
Tell us your idea and we’ll design the right solution for you.
Get in touch