MCP Server for AI Assistants

Clariti's MCP server lets an AI assistant such as Claude Code or Cursor read your Clariti workspace. Ask it "which laptops are missing EDR?", "what are our open critical findings and how do we fix them?" or "why did our posture score drop this month?" and it looks up the answer in your real inventory instead of guessing.

The server is read-only. It has no tools that change anything, and every lookup runs in a read-only database transaction limited to your workspace. The assistant sees what a Read only API key can already see, nothing more.

MCP (the Model Context Protocol) is the open standard AI assistants use to connect to tools and data. If your assistant supports remote MCP servers, it can use Clariti.

Before You Start

You need a Clariti API key with Read only access. An admin creates it:

1. In Clariti, open Settings → API keys.
2. Click Create API key, name it after the assistant that will use it (for example Claude Code - Jane), and choose Read only.
3. Copy the key and store it somewhere safe. It is shown once.

Use a separate key for each person or assistant, so you can revoke one without breaking the others. The MCP server refuses Export only and Read & write keys: an assistant only ever gets read access. See API Keys & Bulk Export for how keys work and when they stop working.

Connect Your Assistant

Claude Code

With your key in the CLARITI_API_KEY environment variable:

claude mcp add --transport http clariti https://api.clariti-ai.xyz/mcp \
  --header "Authorization: Bearer $CLARITI_API_KEY"

Then ask Claude about your environment, for example "Using Clariti, list our critical findings and how to fix each one."

Other MCP clients

Any client that supports remote (Streamable HTTP) MCP servers with a custom header can connect with the settings in the table above.

A client that only runs local MCP servers can reach Clariti through the open-source mcp-remote bridge. Put this in the client's MCP configuration, with your key in place of clk_...:

{
  "mcpServers": {
    "clariti": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://api.clariti-ai.xyz/mcp", "--header", "Authorization:${CLARITI_AUTH}"],
      "env": { "CLARITI_AUTH": "Bearer clk_..." }
    }
  }
}

The configuration file holds your key, so keep it out of source control and shared folders.

Tools

nl_query only runs Clariti's curated templates. A question they don't cover gets a list of what the assistant can ask instead, never a query made up on the fly.

The overall score and the coverage figures are fractions from 0 to 1: 0.82 means 82%. Exposure category scores, exposure_overall and projected_lift are on a 0 to 100 scale.

Each day's score_model_version says which scoring model produced its overall score: 1 is the coverage model, 2 and later the exposure model (3 since 2026-10), which also counts vulnerabilities and configuration. Scores from different models measure different things, so the assistant should not compare them; the posture_trend and posture_biggest_drop templates refuse to and say when the model changed.

What the Assistant Can and Cannot Do

• Can read your workspace's assets, findings, remediation steps and posture history.
• Cannot change anything: no triage, no status changes, no settings, no integrations.
• Cannot see any other workspace. The key decides the workspace, and database row-level security enforces it on every lookup.

What the assistant does with the answers is up to the assistant you choose. Its provider receives the data it reads, the same as any other tool you connect to it. Connect only assistants your organization has approved for security data.

Limits and Errors

The MCP server accepts up to 100 requests a minute per workspace, the same limit as each API endpoint. A single question can take several tool calls.

For assistants and scripts that read documentation directly, a plain-text reference is at docs.clariti-ai.xyz/mcp.md and a site summary at docs.clariti-ai.xyz/llms.txt.