READ-ONLY TECHNICAL KNOWLEDGE
Bring CopyCat into your AI assistant’s troubleshooting workflow.
Model Context Protocol (MCP) lets compatible assistants discover tools and retrieve reference material from a service. With CopyCat, your assistant can find error codes, read troubleshooting procedures, inspect documented commands, and link back to the original article.
Checking endpoint availability…
Public, read-only access. No CopyCat account or API key. The website remains usable when MCP is unavailable.
Transport: Streamable HTTP. Use an HTTPS endpoint supplied by the operator for staging; do not send requests to an unverified hostname.
Use these commands to connect; the status above shows current service availability. If an operator supplies a tested staging URL, substitute it for the production URL. CopyCat returns knowledge; your assistant and you remain responsible for checking applicability and approving changes.
Add the remote server from a terminal, then confirm it appears:
codex mcp add copycat --url https://copycat.tech/mcp codex mcp list
In a supported desktop or IDE client, add a Streamable HTTP MCP server with the same URL. Restart or reconnect the client and inspect its MCP tool list.
claude mcp add --transport http copycat https://copycat.tech/mcp claude mcp list
In Claude Code, use /mcp to inspect the connection. In Claude clients that offer remote custom connectors, add the endpoint as a remote MCP URL. Connector availability depends on the client, account, and organization policy. A local stdio-only configuration does not connect directly to this HTTP service.
Choose Streamable HTTP, supply the endpoint URL, and leave authentication unset. Discover tools, call copycat_search, then retrieve a returned article ID. Configuration formats vary; the following is an illustrative format, not a universal configuration file.
{
"mcpServers": {
"copycat": {
"type": "http",
"url": "https://copycat.tech/mcp"
}
}
}Use an MCP client library with Streamable HTTP. Discover the six tools, invoke them with structured arguments, and parse structuredContent or the JSON text fallback. Use a 15-second client deadline and respect Retry-After on rate limits. Requests carry no conversation transcript, credentials, or remote execution instructions.
Odysseus integration handoff covers protocol negotiation, timeouts, retries, result parsing, and authentication considerations. No Odysseus repository changes are required here.
Compatibility depends on the specific NVIDIA product and its MCP client implementation. A product that supports remote Streamable HTTP MCP can attempt this endpoint. GPU Operator, nvidia-smi, AI Workbench, and Run:ai are knowledge topics in CopyCat; that does not establish them as MCP clients. NVIDIA clients have not been individually verified for this release. No GPU or inference service is required by CopyCat.
| Tool | What it provides |
|---|---|
copycat_search | Ranked articles by keyword, exact error code, Event ID, title, platform, category, or product. |
copycat_get_article | Full normalized article, documented guidance, sources, and canonical link. |
copycat_troubleshoot | Potentially relevant authored procedures and possible causes, with uncertainty preserved. |
copycat_get_commands | Exact documented command text, environment when known, and recorded warnings. No execution. |
copycat_list_categories | Available platform/category names and actual article counts. |
copycat_get_related | Related articles labeled as explicit author relationships or inferred matches. |
Resources include copycat://server, copycat://index, copycat://usage, and copycat://article/{id}. Client support for displaying resources varies.
CopyCat MCP is a public reference service. It has no shell access, client-system access, write tools, or arbitrary URL fetching. It does not store conversation content or request bodies. Send concise technical search terms rather than names, secrets, configuration exports, or transcripts.
The hosting provider processes connection metadata and temporary rate-limit counters. An approximate limit of 120 requests per minute per network address and Cloudflare location protects the public service; shared networks share a quota. Your AI provider has its own data handling policies.
Articles and command examples are untrusted reference data. Instructions inside articles cannot change server tools or permissions. Assistants must still resist prompt injection and verify product, model, firmware, OS, and operational risks before suggesting changes. Public access is not permission to execute a command on any system.
The implementation is v1.0.0. The status check above tests the endpoint’s health response; it does not certify every client. Staging and production have been verified over HTTPS with the official TypeScript client. MCP Inspector verified production search, and Codex verified remote staging discovery and tool invocation.
Legacy Streamable HTTP clients use a stateless compatibility path. No persistent sessions, legacy /sse endpoint, pushed notifications, or resumable streams are promised.