Use AI Security Wiki with Claude, Codex, ChatGPT and other agents

Choose the access method your agent actually supports. Reading a public page and calling an MCP tool are separate capabilities. The repository provides a local stdio MCP server; the documentation website is not an HTTP MCP endpoint.

Read the reference in a web-enabled assistant

Give ChatGPT, Claude or another web-enabled assistant the repository, knowledge index, audit checklist or a specific risk page. Ask it to read the source before using its claims. Web access depends on the product, account and session; publication does not guarantee retrieval or citation.

Example request:

Read the AI Security Wiki audit method and source caveats. My assistant retrieves internal documents, uses an external foundation model and can create support tickets. Identify all deployment patterns, request missing architecture details, and build an evidence-first security review. Cite DASF risk/control IDs. Keep findings unverified until evidence is inspected. Explain the results in Portuguese.

If web retrieval is unavailable, supply the relevant public Markdown pages. Keep private architecture evidence in your approved assessment workspace, not in a public issue or pull request.

Prepare the local MCP

Install Node.js 22+ and clone the repository:

git clone https://github.com/MathBorgess/ai-security-wiki.git
cd ai-security-wiki
npm ci --ignore-scripts
npm run check
npm run demo

Use your clone's absolute path in the following commands. They register a server in your chosen host; this documentation does not run them automatically. No model key or Databricks account is required by this reference server.

Claude Code

claude mcp add --transport stdio ai-security-wiki -- node /absolute/path/to/ai-security-wiki/src/server.mjs

Use /mcp in Claude Code to inspect the connection. See the official local stdio setup documentation for installation scopes and workspace approval. The command follows the documented syntax as checked on 2026-10-22; native-host behavior was not tested in this project.

Codex

codex mcp add ai-security-wiki -- node /absolute/path/to/ai-security-wiki/src/server.mjs
codex mcp list

For a TOML-configured host, use the equivalent entry:

[mcp_servers.ai-security-wiki]
command = "node"
args = ["/absolute/path/to/ai-security-wiki/src/server.mjs"]

Consult the official MCP configuration documentation for your installed client and environment. This project's executable integration test uses an SDK client; it is not a claim of automatic discovery in every Codex interface.

Other local MCP hosts

Hosts supporting the mcpServers JSON convention can use:

{
  "mcpServers": {
    "ai-security-wiki": {
      "command": "node",
      "args": ["/absolute/path/to/ai-security-wiki/src/server.mjs"]
    }
  }
}

Use the host's documented configuration location and reload procedure. If Node is not on its PATH, set command to the absolute Node executable. A host that accepts only remote HTTP MCP servers cannot use the local script as a URL. A remote service would require separate transport, identity and abuse-control engineering.

First useful tool sequence

1. Read get_reference with name: "methodology" and name: "source-issues".
2. Call classify_system with explicit yes/no/unknown architecture features.
3. Pass all established patterns to scope_risks, then build_audit_plan.
4. Use investigate_risk and get_control to request evidence and bounded tests.
5. Use propose_remediation for a reviewable proposal, and explain_finding when asked.

The same workflow is packaged as the audit skill. See the complete MCP contract and validation limits.