Qualien Mcp

qualien-mcp

One MCP server that gives your AI assistant a QA toolkit: Playwright, your test files, GitHub, Jira, Postgres and more, behind a single connection.

[npm] [license: MIT]

qualien-mcp is a composite MCP gateway for SDET / QE work. It aggregates Model Context Protocol servers behind one connection, with a curated catalog of 21 testing-relevant servers, per-tool allow/deny, and read-only-by-default guardrails. Add one server to Claude Code, Claude Desktop, Cursor or VS Code instead of five.

npx qualien-mcp

Out of the box it serves the QE starter pair: Playwright MCP (browser automation, DOM, screenshots) and Filesystem MCP (read/edit project files, page objects, test utils), with zero config. Tools are exposed to your assistant namespaced as playwright__… and filesystem__….

Why a gateway?

Your assistant can already connect to many MCP servers directly — so a gateway earns its place by doing what a pile of servers can't:

• One endpoint to add to any client (Claude Code, Claude Desktop, Cursor, …).
• Curation — enable/disable per server and per tool. Fewer, better-named tools = better tool-selection by the model (dumping 200 tools at it makes it worse).
• One routing/logging surface — every call is logged to stderr with the downstream, tool, and latency.
• Clean lifecycle — spawns each downstream as a child process and tears them all down on exit (no orphaned browser processes).
• Curated for QE — an opinionated SDET toolkit, not a generic proxy.

Use it with your assistant

Claude Code:

claude mcp add qualien -- npx -y qualien-mcp

or in .mcp.json / user settings:

{  "mcpServers": {    "qualien": { "command": "npx", "args": ["-y", "qualien-mcp"] }  }}

Claude Desktop (claude_desktop_config.json) and Cursor (~/.cursor/mcp.json or .cursor/mcp.json): the same shape under mcpServers.

Claude Desktop extension: download qualien-mcp-<version>.mcpb from the latest release and open it. Claude Desktop asks for a project folder: the Filesystem tools are limited to it, and a qualien-mcp.config.json inside it turns on more servers. Also listed on Smithery.

VS Code (.vscode/mcp.json):

{  "servers": {    "qualien": { "command": "npx", "args": ["-y", "qualien-mcp"] }  }}

That single entry gives your assistant every tool from every aggregated server.

Add more servers

Drop a qualien-mcp.config.json in your working directory (or pass --config <path>). Entries merge over the built-in defaults — redefine a key to change it, or set enabled: false to turn a default off. No new release needed.

{  "servers": {    "filesystem": {      "command": "npx",      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/abs/path/to/allow"]    },    "sequential-thinking": {      "command": "npx",      "args": ["-y", "@modelcontextprotocol/server-sequential-thinking"]    },    "chrome-devtools": {      "command": "npx",      "args": ["-y", "chrome-devtools-mcp@latest"]    },    "playwright": {      "command": "npx",      "args": ["-y", "@playwright/mcp@latest"],      "tools": { "deny": ["browser_close"] }    }  }}

See qualien-mcp.config.example.json for the full shape (command, args, env, enabled, tools.allow / tools.deny).

Built-in catalog

qualien-mcp ships a curated catalog of 21 known servers (each verified against npm), so you enable one by key with just your secrets — no need to remember packages or commands:

{  "servers": {    "chrome-devtools": { "enabled": true },    "postgres":  { "enabled": true, "env": { "DATABASE_URI": "postgres://…" } },    "slack":     { "enabled": true, "env": { "SLACK_MCP_XOXP_TOKEN": "xoxp-…" } }  }}

Run npx qualien-mcp catalog to list them all with what each needs. Current catalog:

use mounts a catalog server under a different key (e.g. two databases): { "db-prod": { "use": "postgres", "enabled": true, "env": {…} } }. Anything not in the catalog you still define in full (command/args or type/url).

Remote & OAuth servers (e.g. GitHub)

Downstreams can be remote (Streamable HTTP) as well as local. A remote server is { "type": "http", "url": "…" }, and if it needs OAuth, each user logs in with their own account — tokens are stored per user at ~/.qualien-mcp/credentials.json (0600), on that user's own machine, and are never bundled, shared, or sent anywhere by qualien-mcp.

GitHub's hosted MCP is the reference case, and it needs no setup beyond one command:

1. Turn it on in qualien-mcp.config.json:
   json

   { "servers": { "github": { "enabled": true } } }

2. Authorize once:
   bash

   npx qualien-mcp login github

login prints a short code and a URL; you approve it on github.com and you're done. There is no OAuth App to register and no client secret anywhere.

To use your own GitHub OAuth App instead of the shipped one, set clientId (and enable the device flow on it under Settings → Developer settings → OAuth Apps):

{  "servers": {    "github": { "clientId": "<your client id>", "enabled": true }  }}

If a remote server isn't logged in yet, the gateway skips it with a hint (run: npx qualien-mcp login github) and still serves everything else — it never blocks startup. Servers that support dynamic client registration need no clientId at all.

Safe by default

qualien-mcp enforces guardrails centrally, before forwarding a call — so they hold no matter what the downstream permits, and composite tools can't bypass them either:

• Databases are read-only — postgres/mysql calls containing write/DDL SQL (INSERT/UPDATE/DELETE/DROP/…) are refused. Opt in with { "postgres": { "readOnly": false } }.
• Filesystem roots — { "filesystem": { "roots": ["./src", "./tests"] } } refuses any call whose path argument escapes those directories.
• Destructive infra tools blocked — docker/kubernetes tools whose name looks destructive (delete/remove/prune/kill/…) are refused unless { "allowDestructive": true }.
• Plus per-server tools.deny / tools.allow curation.

Honest scope: these stop an LLM from accidentally doing damage. The SQL check is keyword-based (heuristic), not a parser — it is not adversarial sandboxing. Real isolation needs the downstream's own permissions or a container.

Composite QE tools

Beyond passthrough, qualien-mcp ships tools it implements itself (namespace qe__) that orchestrate several downstreams in one call — the QE payoff of a gateway.

qe__verify_api_vs_db — end-to-end API↔DB consistency in one call. Give it two sub-calls (namespaced tools you can discover via tools/list); it runs both and diffs the payloads:

{  "api": { "tool": "openapi__getUser", "arguments": { "id": 1 } },  "db":  { "tool": "postgres__query", "arguments": { "sql": "select id, name from users where id = 1" } },  "match": "subset"        // every field the DB returns must match the API (default)}// → { "match": false, "differences": [ { "path": "name", "api": "Ann", "db": "Bob" } ], … }

Composite tools appear only when their required downstreams are connected (qe__verify_api_vs_db needs a database server). More to come (repro-from-Jira, page-object-from-URL, flaky triage).

How it works

In one process, qualien-mcp is both an MCP server to your assistant and an MCP client to each downstream:

assistant ⇄ qualien-mcp ⇄ playwright-mcp                         ⇄ filesystem-mcp                         ⇄ …

• tools/list fans out to every downstream, filters by your curation rules, and namespaces each as <server>__<tool>.
• tools/call routes by that prefix to the owning server and passes the response straight through.
• Resources and prompts are aggregated the same way when a downstream provides them.
• A downstream that fails to start is logged and skipped — the gateway still serves the rest.
• Logs go to stderr (stdout is the protocol); your host surfaces them in its MCP logs.

Roadmap

• ✅ Playwright + Filesystem, namespacing, curation, logging, clean lifecycle
• ✅ Config-driven catalog of 21 servers, remote/OAuth servers (GitHub via device flow), safe-by-default guardrails, qe__verify_api_vs_db
• Next: more composite QE tools (repro-from-Jira, page-object-from-URL, flaky triage)

Feedback

Found a bug, want a server added to the catalog, or have an idea? Email [email protected] or open an issue. qualien-mcp is part of qualien.ai, which also builds AI agents that test web apps (Olympus) and native Android apps (Pulse) from plain English.

License

MIT