GenClass MCP sunucusu
Commit 12 Kas 2026Lets AI agents observe web app internals, identify race conditions, replay network orders, and generate Playwright tests to prove bug fixes.
Ayrıca bilinen adı genclass-dev/genclass-plugin
- Yıldızlar
- 0
- Araçlar
- —
- İndirmeler, 30 gün
- —
- En son sürüm
- —
- Son commit
- 12 Kas 2026
- Lisans
- Apache-2.0
GenClass MCP nedir?
GenClass bir MCP sunucusudur. Lets AI agents observe web app internals, identify race conditions, replay network orders, and generate Playwright tests to prove bug fixes.
Özet, projenin README dosyasından oluşturuldu
GenClass MCP sunucusu hakkında
Developers using AI agents like Claude Code, Cursor, or Gemini CLI can use this plugin to observe the internal workings of web applications running locally. It allows an agent to monitor every click, network request, and response order to identify complex frontend issues. By connecting to a development build via a local bridge on 127.0.0.1:47357, the server provides deep visibility into application state and uncaught errors.
The system includes nine specialized tools such as genclass_scan, genclass_replay, and genclass_findings to help agents diagnose problems. It can simulate different network response orders to uncover race conditions and replay specific sequences to verify bug fixes. Once a bug is identified, the agent can generate Playwright tests to ensure the issue does not return. This setup runs entirely on your machine, keeping findings in your local project directory unless you explicitly enable external analytics or hosted scanning.
Neler yapabilirsiniz
- 1Identify intermittent race conditionsAsk the agent to find race conditions by re-running a page with the network answering in different orders.
- 2Verify bug fixes with replaysHave the agent replay one exact network order to prove that a specific fix works.
- 3Automate regression testingInstruct the agent to write a Playwright test that fails while the bug is still present.
- 4Monitor application state changesUse the agent to see what each network response writes into the application state.
Projenin README dosyasından alınmıştır. Her kullanım durumu, oradaki bir alıntı ile kontrol edilir.
GenClass plugin
For Claude Code, Codex, Cursor, Gemini CLI and other agents that read plugins or skills. GenClass lets the agent see what happens inside the pages of your web app running on your machine (each click, request and response, the order the responses were delivered, what each response wrote into the app's state, uncaught errors), find race conditions by re-running a page with the network answering in different orders, replay one exact order to prove a fix, and write a Playwright test that fails while the bug is there.
Install
Claude Code (the server, a prompt hook that brings new findings to each prompt, and the genclass skill):
/plugin marketplace add genclass-dev/genclass-plugin
/plugin install genclass@genclass
Codex (the server and the skill): codex plugin marketplace add genclass-dev/genclass-plugin, then
codex plugin add genclass@genclass. Gemini CLI (the same): gemini extensions install https://github.com/genclass-dev/genclass-plugin.
The skill alone, in any agent the skills CLI knows: npx skills add genclass-dev/genclass-plugin; for GitHub Copilot,
gh skill install genclass-dev/genclass-plugin genclass.
Then connect your app's development build and load a page of the dev server. genclass_setup prints the install line
(npm install @genclass/runtime@0.3.0-beta.6, the version this release was tested with; the page side needs 0.3.0-beta.1
or later) and the lines for your framework, behind a development check. See the
package README.
What the agent asks first
The skill and the server's instructions tell the agent to show you each of these and wait for your yes: running the
install or adding the lines genclass_setup prints; npx playwright install chromium and npm i -D playwright; genclass_scan with
hosted: true; writing a fix or a test when you only asked what broke. These are instructions to the agent, not locks
in the code; your client's own permission prompts still apply.
What it installs
| part | file | what it does |
|---|---|---|
MCP server genclass |
.mcp.json |
runs the GenClass MCP server, @genclass/mcp@0.2.2 (vendored as server/cli.mjs in the plugin repository and run with node; the copy inside GenClass runs npx -y @genclass/mcp@0.2.2): nine tools (genclass_status, genclass_findings, genclass_explain, genclass_explore, genclass_replay, genclass_analytics, genclass_setup, genclass_scan, genclass_clear) and a bridge on 127.0.0.1:47357 that your app's pages connect to |
UserPromptSubmit hook |
hooks/hooks.json |
runs the same server's hook command before each prompt: up to five findings recorded since the last prompt are added to the prompt as context |
skill genclass |
skills/genclass/SKILL.md |
the procedure once Claude reaches for the tools (the explore arguments and Playwright pairing, how to read a report, what to say when done, what to do when the tools are missing); its description carries the triggers (an intermittent frontend bug; after changing code that fetches, submits, autosaves, polls or debounces; before calling a UI change done), as do the server's instructions. Without the plugin, npx -y @genclass/mcp@0.2.2 init writes the same file to .claude/skills/genclass/SKILL.md |
In the plugin repository the server is vendored: server/cli.mjs is the published @genclass/mcp@0.2.2 build (its
dist/cli.js, byte for byte, fetched from npm by the daily sync), server/genclass-mcp.mjs is the package's command
importing it, and .mcp.json and the hook run them with node from ${CLAUDE_PLUGIN_ROOT}, so installing the plugin
installs nothing and runs the exact release it was reviewed with (the one download is genclass_scan's, below); there is no package.json or lockfile. The copy of
this folder inside GenClass runs npx -y @genclass/mcp@0.2.2 instead (nothing built is committed there).
plugin.json and mcp.json at this folder's root are the same server in the
Agent Plugins format, for Cursor, Codex and other clients that read it (server and skill,
no hook); that format has no plugin-root variable, so they keep npx.
What the hook sends, and where
Only to this machine. It reads the bridge's token from ~/.genclass/bridge-47357.token (mode 0600), asks the bridge
at http://127.0.0.1:47357 for findings newer than its cursor, keeps that cursor in .genclass/agent/hook-cursor in
the project, and prints the new findings' ids, kinds, pages and titles for Claude Code to add to the prompt. Page
strings are quoted and labelled as recorded data, not instructions. It sends nothing else anywhere, prints nothing when
no bridge is running or nothing is new, and gives up after 0.8 seconds if the bridge does not answer. The first prompt
of a session only sets the cursor (genclass_findings has the backlog).
The server listens on 127.0.0.1 only and accepts pages from this machine (localhost, 127.0.0.1, *.localhost)
or origins you list in GENCLASS_AGENT_ORIGINS. Nothing leaves your machine unless you or your agent ask
for it: genclass_analytics {source: "production"} reads your own GenClass dashboard when you set
GENCLASS_DASHBOARD to its link; genclass_scan with hosted: true sends the URL to the genclass.dev page check
(without it, the page loads in a local headless browser); and GENCLASS_REPORT, off by default, mirrors finding ids,
kinds, page paths, replay verdicts and counts to that dashboard, never titles or page text. When the project has no
@genclass/runtime, genclass_scan downloads the version this release pins, @genclass/runtime@0.3.0-beta.6, from npm
through npx.
Details: AGENT.md section 6.
Privacy
By default the plugin collects nothing and sends nothing to GenClass or anyone else; the exceptions you can turn on are
listed above. Findings stay in .genclass/agent/ in your
project (gitignored by the server) and in the bridge's memory.
The lines genclass_setup prints start the runtime in development builds only, and a production build drops them. On a
local or private address the runtime sends no diagnostics (it does fetch its on-device model from jsDelivr). Running it
in production is a separate choice this plugin does not need: outside the check GenClass.init() runs for every visitor
and sends anonymous diagnostics from public hosts by default, which GenClass.init({ telemetry: false }) turns off
(PRIVACY.md).
Turn it off
- Everything:
/plugin disable genclass@genclass(or/plugin uninstall genclass@genclass). - Only the hook: Claude Code has no switch for one plugin hook; uninstall the plugin and add the server alone
(
claude mcp add genclass -- npx -y @genclass/mcp@0.2.2), or set"disableAllHooks": truein your settings, which turns off every hook. - In the app: remove the
connectAgent(rt)block; the page then never talks to the bridge.
Where this comes from
This folder is packages/mcp/plugin in genclass-dev/GenClass.
The repository genclass-dev/genclass-plugin is this folder at a
repository root with the marketplace files and the vendored server, assembled by
node packages/mcp/scripts/plugin-repo.mjs and re-synced daily, so issues and pull requests go to GenClass.
License
Apache-2.0. Source: packages/mcp.
GenClass MCP soruları
What can GenClass help an AI agent do?
It allows agents to see clicks, requests, responses, and app state changes to find race conditions, replay specific network orders, and write failing Playwright tests.
Is my data sent to external servers?
By default, the plugin collects and sends nothing; findings stay local to your machine and project. You must explicitly enable specific features like analytics or hosted scans.
How do I install it for Claude Code?
You can install it using the command /plugin marketplace add genclass-dev/genclass-plugin followed by /plugin install genclass@genclass.