MCP server: the catalogue in your AI assistant

The “contexte” MCP server gives your AI assistant direct access to the catalogue: it searches resources, reads a dataset, fills a prompt, with nothing for you to download or copy. It reads the public API: no account, no key.

Online, nothing to install

For ChatGPT, Claude on the web and mobile, and any assistant that accepts an MCP server URL: add a connector with this URL, no authentication.

https://contextetech.com/mcp

ChatGPT : Settings → Apps & Connectors → Advanced settings → developer mode, then “Create” a connector with the URL above.
Claude (web, mobile) : Settings → Connectors → “Add custom connector” with the URL above.

Same tools, except saving a file, replaced by get_download_link (dataset download link). The search and fetch tools follow the ChatGPT connector format.

On your computer (Claude, Cursor, VS Code)

Requirement

uv, the tool that starts the server (uvx downloads the “contexte” Python package and runs it):

curl -LsSf https://astral.sh/uv/install.sh | sh     # Mac, Linux
brew install uv                                      # Mac (Homebrew)
pip install uv                                       # or with Python

Installation

Add this configuration to your assistant, then restart it.

Claude (app)

Settings → Developer → Edit config (claude_desktop_config.json file):

{ "mcpServers": { "contexte": { "command": "uvx", "args": ["contexte@latest", "mcp"] } } }

Claude Code

In a terminal:

claude mcp add contexte -- uvx contexte@latest mcp

Cursor

File ~/.cursor/mcp.json:

{ "mcpServers": { "contexte": { "command": "uvx", "args": ["contexte@latest", "mcp"] } } }

VS Code

File .vscode/mcp.json in the project:

{ "servers": { "contexte": { "type": "stdio", "command": "uvx", "args": ["contexte@latest", "mcp"] } } }

If the assistant cannot find uvx, replace "uvx" with its full path (for example /opt/homebrew/bin/uvx on a Mac, given by the command which uvx).

Tools

Tool	What it does
`search_resources(query, type?, language?)`	Searches the catalogue: short code, title, type, language, licence, number of rows and link of each resource.
`load_resource(code, rows?)`	Loads a resource by its short code (end of its URL, for example 13n9ym0): first rows of a dataset, template of a prompt, instruction and examples of a context.
`fill_prompt(code, variables)`	Fills a prompt's {{variables}} and returns the ready-to-use text.
`download_dataset(code, format?, path?)`	Saves the whole dataset to a file: original JSONL, messages (OpenAI), ShareGPT, Alpaca or CSV. Never overwrites an existing file.
`estimate_cost(code)`	Estimated cost in euros: tokens, cost per call or per pass, batch discount, cheapest model, fine-tuning on a rented GPU.

All tools are read-only, except download_dataset, which writes a file. “@latest”: the assistant takes the latest server version at each start.

Publish your resources from the assistant

With an API token, the assistant creates and updates your resources: describe the resource, it writes and saves it. A new resource is a draft only you can see: check it on its page, then ask the assistant to publish it.

1. Create a token in Profile → API tokens. 2. Add it to the online server in your assistant:

Claude Code

In a terminal:

claude mcp add --transport http contexte https://contextetech.com/mcp --header "Authorization: Bearer ctx_…"

Cursor

File ~/.cursor/mcp.json:

{ "mcpServers": { "contexte": { "url": "https://contextetech.com/mcp", "headers": { "Authorization": "Bearer ctx_…" } } } }

VS Code

File .vscode/mcp.json in the project:

{ "servers": { "contexte": { "type": "http", "url": "https://contextetech.com/mcp", "headers": { "Authorization": "Bearer ctx_…" } } } }

Tool	What it does
`publish_resource(type, name, title?, description?, language?, license?, content?, draft?)`	Creates or updates one of your resources (draft by default, draft=false to publish). On an update, fields left out keep their value.
`list_my_resources()`	Lists your resources, drafts included: code, title, type, status and link.

Same checks as the form: validation, content safety, at most 20 new resources a day, moderation. The token only gives access to your resources, never to your account. Without a token, these two tools do not appear.

Example requests

“Find an English dataset to get the HS customs code of a product.”
“Load 13n9ym0 and show me three examples.”
“Save sw3sz9 as ShareGPT in my project.”
“How much does one pass of sw3sz9 cost with Mistral Large?”

Security

Without a token, no change to the catalogue: the server publishes, edits and deletes nothing online; only download_dataset writes a file on your computer, never overwriting one.
No key and no personal data: it only sends your search to the public API.
Content written by members is flagged to the assistant as data, not instructions; resources are checked on publication (hidden instructions, dangerous commands, secrets).

Official registry

The server is listed in the MCP registry as io.github.contexte-tech/contexte (PyPI package “contexte”).