veridicusscan-mcp-analyst（veridicusscan MCP analyst 搜索）

VeridicusScan MCP Analyst

Use this skill only for the VeridicusScan MCP surface, not for changing the app code itself.

VeridicusScan is a local-first scanner and runtime-defense tool. This skill is for analyst tasks such as scanning websites, files, prompts, job-application artifacts, and agent-runtime flows through the MCP bridge.

Preconditions

• Confirm a VeridicusScan MCP server is available in the client.
• If it is not available, say so briefly and ask the user to connect the local bridge first.
• Prefer the MCP server over shelling out to app internals when both can do the task.
• Expect one active MCP session at a time. If open_session returns session_limit_reached, tell the user another active session is still open.

High-value use cases

• Scan a public website or candidate portfolio URL before an AI agent reads it.
• Scan a local PDF, DOCX, image, or exported text artifact before model handoff.
• Triage prompt snippets or extracted page text with scan_text.
• Validate agent-memory and tool-approval flows with the runtime-defense methods.

Core workflow

1. Start with health or list_methods if availability is unclear.
2. Open a session with open_session.
3. Run the smallest relevant scan method:

- scan_url for live public HTTPS websites - scan_file for local files - scan_text for prompts, snippets, and extracted content

1. Pull the report or scan result details the user actually needs.

- If scan_file returns default_context_mode = "sanitized_only", prefer safe_context for downstream use and make clear that report surfaces are redacted by design. - If scan_url returns non_public_network_url, explain that VeridicusScan intentionally blocks loopback, private-network, .local, .localhost, and resolved internal targets.

1. Summarize:

- risk band - risk score - default context mode when present - findings count - top findings with short evidence summaries - coverage limits or partial-scan notes

1. Close the session when done unless the user is actively continuing a multi-step analysis.

Reporting rules

• Be explicit about whether a result is a likely true positive, likely false positive, or uncertain.
• If the scan is partial, explain exactly what was not covered and why that matters.
• If a result is redacted or sanitized_only, say that explicitly instead of implying raw evidence is available.
• Distinguish structural signals from semantic injection signals.
• For benign sites, do not overclaim. Say when a hit looks like tracking, accessibility, anti-bot, or app-shell markup rather than malicious prompt injection.
• Include exact MCP error codes when they change the user outcome, for example non_public_network_url or session_limit_reached.

Runtime-defense workflow

Use these methods when the user is evaluating agent safety rather than content scanning:

• ingest_memory for A1 memory ingestion
• retrieve_memory for A2 retrieval validation
• selective_disclosure and evaluate_selective_disclosure for disclosure quality and privacy checks
• scope_tools before planning or execution
• guard_plan before approving a plan
• gate_action before approving a specific tool action

Always preserve the returned tool scope and pass the authoritative scope back into guard_plan and gate_action. Do not invent or forge scope values.

Recommended ordering for tool-governance tasks:

1. open_session
2. scope_tools
3. guard_plan
4. gate_action
5. close_session

Output style

• Keep summaries short and operational.
• Put findings first.
• Include exact method names when explaining how a result was obtained.
• If the user asks for verification, say which MCP method(s) you used.
• Prefer language that is useful to operators, for example security, IT, recruiting, or compliance teams, instead of purely academic scanner jargon.

References

• Read references/mcp-methods.md for the method map and sequencing guidance.
• Public workflow example for recruiting and job-application intake:

AI job application screening

• Public technical workflow example for MCP sessions and runtime guardrails:

Local MCP automation for AI agents