Qlane agent plugin

Installs Qlane MCP into your AI coding editor, so your agent can reach your Qlane projects, test targets and what Qlane found on your pull requests without leaving the editor.

Which region?

Qlane runs two independent regional stacks. The plugin registers one server for each, and you connect the one hosting your organization:

Authenticate that one and leave the other alone. An unconnected server contributes none of its own tools — your editor shows it with two sign-in helpers and nothing else — so it costs you nothing and is not an error state. If you connect the wrong one, every call returns a 403 naming the correct host — nothing is lost, and the fix is to connect the other server.

Both are plain static URLs, so this works identically on every editor. Nothing is configured, substituted, or typed.

Install

Qlane MCP requires a Qlane account. Your editor prompts you to sign in the first time it calls the server.

Claude Code

claude plugin marketplace add qlaneai/agent-plugin
claude plugin install qlane@qlane-plugin

Then run /mcp and authenticate qlane-eu or qlane-us — whichever hosts your organization. There is no configuration step and nothing to type.

VS Code and GitHub Copilot

Add this repo to chat.plugins.marketplaces:

{
  "chat.plugins.marketplaces": ["qlaneai/agent-plugin"]
}

If nothing appears after adding it, check that chat.plugins.enabled is true.

Cursor

In Cursor's Import Marketplace dialog, enter the full repository URL, https://github.com/qlaneai/agent-plugin — Cursor rejects the qlaneai/agent-plugin shorthand there. Then install qlane and authenticate the server for your region.

Cursor also loads .cursor-plugin/plugin.json, which gives the same two servers Qlane's pre-registered Cursor client. So Cursor signs in as Qlane's verified Cursor client, and an organization's default client policy admits it.

Connected a Qlane server with an earlier version of this plugin? After the update, Cursor should ask you to authenticate the server again; do, and it signs in as the verified client. If it keeps working without asking, sign out of the server in Cursor (or remove its stored authentication) and authenticate again.

Cursor can also import this plugin from a Claude Code installation. An imported copy stays at the version Claude Code has installed, and only a version that carries .cursor-plugin/plugin.json signs in as the verified client, so update it there first: claude plugin update qlane@qlane-plugin.

Kiro

Kiro reads the root plugin.json — the portable manifest — and loads this plugin without changes. Add the repository as a marketplace, install qlane, then authenticate the server for your region.

Codex

Codex splits this across two places: you register the marketplace on the CLI, but you install from it in the ChatGPT desktop app. There is no CLI install command.

codex plugin marketplace add qlaneai/agent-plugin

Then open the Plugins Directory in the ChatGPT desktop app and install qlane from it. Codex prompts you to sign in the first time it calls the server.

Opening this repository in Claude Code

You will see qlane-eu and qlane-us listed as Pending approval. That is expected and harmless: Claude Code reads a .mcp.json at a repository root as project-scoped MCP configuration, so this repo's .mcp.json — the plugin's own manifest — is being read as if it were yours. Decline it, and install the plugin with the commands above instead.

Without the plugin

Any MCP client can point at the endpoint directly. For Claude Code:

claude mcp add --transport http --scope user qlane https://mcp-eu.qlane.ai/api/mcp

Swap the host for mcp-us.qlane.ai if your organization is US-hosted.

What it can do

Set up a test target for a repository, read your projects and test targets, and see what Qlane found when it tested your pull requests: each run's verdicts, the defects it saw and the evidence behind them. Three skills put that to work:

• /qlane:setup-env takes a repository to a test target whose configuration check passed. It finds the Qlane project the repository belongs to, or creates one, prepares the repository from Qlane's published setup pages (docs.qlane.ai/connect), creates the test target, and has you enter its variables and test credentials on a Qlane page — they never pass through the editor. It asks before every change to your repository or to Qlane's configuration.
• /qlane:bugs shows what Qlane found on a pull request or a recent run — defects, failing and blocked cases, the steps to reproduce them and their evidence — then looks for the likely cause in your repository and proposes a fix. It only reads from Qlane, and it changes your code only when you say yes.
• /qlane:test-local-changes asks Qlane to plan tests for a local diff before a pull request exists, then exercises them against your running app.

Development

npm ci
npm run check

check runs Prettier, markdownlint, JSON Schema validation of the portable manifests, and the unit tests. It makes no network calls, so it works offline and in a sandboxed CI runner.

npm run format      # apply Prettier
npm run lint:md:fix # apply markdownlint's safe fixes

Two checks need the Claude Code CLI and so are not part of check, which has to run on a runner that does not have it:

claude plugin validate --strict .   # manifest shapes; --strict makes warnings fatal
claude plugin tag --dry-run         # what the next release tag would be

The version lives in plugin.json only. The marketplace entry deliberately declares no version of its own: plugin.json wins at install time, so a second copy could only ever agree or be silently ignored.

No manifest may contain a placeholder. Both regional URLs are literal strings in every manifest. ${user_config.*} resolves only in the Claude Code CLI — the portable Agent Plugins schema cannot even express userConfig (additionalProperties: false), and claude.ai treats an MCP url as an opaque connector identity, so a templated host reaches registration verbatim and fails to resolve. A test enforces this.

License

MIT.