Quilr native Python MCP connector

Connect approved local MCP tools to your organization's Quilr gateway. Python MCPs run in Python; Node MCPs run in Node when Node is installed. The connector itself requires only Python 3.12 or newer on macOS, Linux or Windows.

Use the exact version and gateway supplied by your administrator. Once this release is published on PyPI, the one-command setup is:

uvx --from 'quilr-local-mcp-runtime==0.4.1' quilr-mcp connect --gateway 'https://YOUR-GATEWAY' --console 'https://YOUR-CONSOLE' --organization 'YOUR-ORGANIZATION-UUID' --client cursor

Alternatively, install persistently with pipx install 'quilr-local-mcp-runtime==0.4.1' or install the reviewed wheel into your own virtual environment with python -m pip install /path/to/reviewed.whl, then run the same quilr-mcp connect command. A wheel can be installed directly before the release is published. Never run the connector as root.

The command configures the selected AI client and opens browser sign-in. Approve your computer, then open or restart your AI app. Supported clients are Cursor, Claude Code, Claude Desktop on macOS, and VS Code with the code command installed. --client other prints the configuration for another stdio MCP host.

The installation stores a stable launcher and its installed dependencies under your private connection directory; AI-client startup does not depend on keeping the temporary uvx cache. It still needs the OS Python used to connect. Reconnect after replacing that interpreter.

Ask your AI app to check your Quilr connection and list the permitted tools. quilr_connect opens browser approval when needed; quilr_connection_status checks sign-in and package preparation. Credentials never belong in chat. Packages install lazily from reviewed, signed definitions. A Preparing result means the requested tool has not executed; check status and retry once ready.

Python dependencies use a private virtual environment and hash-locked PyPI wheels. Node dependencies use an npm lockfile with integrity checks and lifecycle scripts disabled. Installing a connector through PyPI does not convert Node MCPs into Python, and installing through npm does not convert Python MCPs into Node.

Tool requests and results pass through the existing gateway policy path. The connector does not grant access beyond your organization's policy, intercept built-in AI-client tools, or provide an OS sandbox. Each MCP runs as your OS user. For tools with internal agents, gateway enforcement covers their exposed MCP calls, not every internal browser or model operation.

Use the printed connection directory for advanced commands:

quilr-mcp status --state /absolute/connection-directory
quilr-mcp list --state /absolute/connection-directory
quilr-mcp doctor --state /absolute/connection-directory --package PACKAGE_ID
quilr-mcp disconnect --state /absolute/connection-directory

Disconnect revokes this connection and removes its local credential while preserving package and workspace files. Optional connect --workspace, --python and --node select local paths. Provider environment variables stay local and are forwarded only when declared in the approved package.

Python and npm are alternative connector installations for the same gateway. Use one connector per AI-app connection to avoid duplicate tool registrations. They use the same enrollment and policy protocol.

One gateway, multiple consoles

Use the command copied from the console you signed into. Version 0.3.1 adds --console https://YOUR-CONSOLE: the connector opens that origin's /self-service?setup=local-mcp page, remembers it for future sign-ins, and still sends enrollment exchanges and tool requests only to --gateway. Two consoles can share a gateway without overwriting each other's sign-in destination. The organization remains bound to the pairing and verified during exchange.

An explicit console replaces any old gateway-wide setup URL for this connection. Running connect again without --console preserves the saved choice. Supplying a new console updates it without creating a second connection for the same gateway and organization. Restart the AI app after reconnecting to load the new launcher. Commands without a saved or explicit console keep the gateway's fallback URL.

The console must be an HTTPS origin, without a path, query, fragment or credentials. Local development can use --console http://localhost:3000 with --allow-loopback-http; this never allows HTTP on a remote console. The browser receives only the short-lived user code in a URL fragment, not the verifier or connector token. Approve only connections you initiated.