Ethicore Guardian plugin
A Claude Code plugin that screens prompts, tool calls and documents for injection and leaks, returning BLOCK, CHALLENGE or ALLOW
- 123 ★ on the repository
- Last commit
Extended access
Why this level
- Texts, call arguments and documents go to the Guardian cloud API; a paid key is required.
- An ALLOW verdict does not guarantee the action is safe.
Install
- Works in
Manual install
Add the ethicore-guardian plugin from the repository's `plugins/ethicore-guardian` directory through `/plugin` and set `ETHICORE_API_KEY` in the settings prompt. Do not print the key or store it in the repository.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Texts, call arguments and documents go to the Guardian cloud API; a paid key is required.
- An ALLOW verdict does not guarantee the action is safe.
This is third-party code. Review the repository files before installing.
What it does
Ethicore Guardian is a plugin for Claude with the `guardian-mcp` MCP server and the `ai-threat-screening` skill. Four tools check threats before the agent acts: `analyze` for a prompt, document or web text, `scan_tool_call` for a tool call the model chose, `scan_tool_output` for a tool or RAG result before it re-enters context and `scan_documents` for PDF, DOCX, PPTX, XLSX, RTF and TXT. Each call returns an ALLOW, CHALLENGE or BLOCK verdict with threat categories and reasoning. Analysis runs on the Guardian cloud API through pattern, semantic, behavioural and ML layers across more than 180 categories: prompt injection, jailbreaks, data and secret exfiltration, tool-call abuse and RAG and memory poisoning. The skill tells Claude when to call each tool.
Who it is for. Teams giving Claude Code access to web pages, documents and tools who want untrusted input screened before actions run.
Good fit when
- The agent reads a web page or document from an outside source and an injection may hide in it.
- Before a shell command or a file write you want the model's choice checked.
- An API or RAG result returns into context and must be checked for hidden instructions.
- Uploaded PDFs and office files are scanned before processing.
Not a fit when
- You have no paid key: the plugin requires a Guardian Pro-plan or higher key and free keys are refused.
- Data cannot go to an external API: texts and documents go to the Guardian cloud.
- You need fully local screening: a standalone Python SDK with a separate license exists for that.
Example request
Screen this PDF and the page text with analyze, then call scan_tool_call for the command you are about to run.Limitations
The plugin is paid: a Guardian API key on the Pro plan or higher is required and the free community key is refused. Texts, arguments and documents being screened go to the Guardian cloud API over TLS and, by the authors' statement, are analysed transiently and not kept for training. The plugin carries a proprietary license while the SDK itself is under BSL 1.1. The server starts through `uv` with pinned versions from `uv.lock`, so `uv` must be installed. The tools only read and return a verdict. The latest commit is from November 2026.
How to disable. Disable or remove the plugin with `/plugin` in Claude Code and revoke the API key on the Guardian portal if it is no longer needed.
FAQ
Which tools does the plugin provide?
Four: `analyze`, `scan_tool_call`, `scan_tool_output` and `scan_documents`. Each returns an ALLOW, CHALLENGE or BLOCK verdict.
What happens on a CHALLENGE verdict?
Pause and confirm the action with the user. BLOCK means do not take the action and explain what was flagged.
Is a paid key required?
Yes, the plugin requires a Guardian key on the Pro plan or higher. Free access remains through the API and SDK directly.
Where is the key stored?
The key is declared a sensitive setting: Claude Code stores it in the OS secure credential store and passes it only to the server process through an environment variable.
How does the MCP server start?
With `uv run --frozen --project ${CLAUDE_PLUGIN_ROOT} guardian-mcp` over stdio. The `--frozen` flag forbids re-resolving and installs versions from `uv.lock` with hash checks.
Discussion
No questions or reviews yet.
Related
PluginsSecurity hook for coding agents: scores each tool call with session context and flags injections in results and skills
Agent Governance Toolkit
Agent Governance Toolkit (Claude Code plugin)
A Microsoft plugin for Claude Code whose hooks inspect prompts and tool calls against a policy and block dangerous commands.
A Claude Code plugin that classifies each tool call and the agent's reasoning, then blocks risky ones or asks you first.
A plugin and approval server for Claude Code and Codex: a risky tool call waits for your decision before it runs
AgentGuard checks a coding agent's actions before they run and scans third-party skills for secrets, backdoors and injections.
A plugin that holds dangerous tool calls in Claude Code for a human's confirmation and, optionally, a signed receipt.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| Ethicore Guardian (this card) | no | 123 | |
| AgentGuard | no | 494 | a protection layer for agents from another vendor; Ethicore Guardian gives four MCP screening tools with a verdict |
| HOL Guard | yes | 1,094 | checks before installing skills and servers; Ethicore Guardian screens prompts, calls and results at runtime |
README in short
The repository holds a Python SDK for LLM apps and a separate plugin directory for Claude. The main README is about the SDK and protecting the whole agent loop: input, output, tool calls and tool results.