Arkheia
MCP

Arkheia · MCP Trust Server

Add a trust layer
to your AI agents

Install Arkheia once. Claude, Codex, Gemini, Grok, and local agents can verify risky outputs and leave an audit trail as they work.

Install FreeView Client Setup

§ 00 · Self-Serve Trust Layer

Useful before anyone buys the platform

The MCP server is the small installable wedge: one developer can add runtime verification to their AI work, then share receipts and invite the team once the signal proves useful.

01

Install Arkheia

Add the MCP server to the AI tools you already use. Start with Claude Code or Claude Desktop, then extend to Codex, Gemini, Grok, or local agents.

02

Verify agent output

Your agent can call arkheia_verify on risky responses and get a structured verdict: risk, confidence, flags, and detection ID.

03

Keep the receipt

Each detection can become evidence: what was checked, what fired, and what the workflow should do next.

§ 01 · What It Is

Detection as a native tool

The Arkheia MCP Trust Server exposes detection as two native MCP tools over stdio transport. Any MCP-compatible host — Claude Code, Cursor, or your own agent framework — can call them per-invocation, with no HTTP proxy in the request path.

Tool 1 — arkheia_verify

arkheia_verify(
  prompt:     string,
  response:   string,
  model_id:   string,
  session_id? string   // optional
)

Submits a prompt/response pair for detection. Returns a structured verdict.

{
  "risk_level":   "HIGH",
  "confidence":   0.84,
  "flags":        ["signal_anomaly"],
  "detection_id": "det_abc123"
}

Tool 2 — arkheia_audit_log

arkheia_audit_log(
  session_id: string,
  limit?:     number   // default 20
)

Returns recent detection events for a session — for pattern analysis, audit reporting, or agent self-review.

{
  "events": [
    {
      "detection_id": "det_abc123",
      "risk_level":   "HIGH",
      "model_id":     "claude-opus-4-6",
      "timestamp":    "2026-02-28T..."
    }
  ]
}

§ 02 · Claude Code & Cursor Setup

No endpoint changes. Detection at the tool layer.

Claude Code and Cursor users don't need to change ANTHROPIC_BASE_URL or any API endpoint. The MCP Trust Server integrates at the tool layer — no proxy in the request path.

Claude Desktop — mcpServers config

{
  "mcpServers": {
    "arkheia": {
      "command": "npx",
      "args": ["-y", "@arkheia/mcp-server"],
      "env": {
        "ARKHEIA_API_KEY": "ak_live_your_key_here"
      }
    }
  }
}

01

Install once

Add the Arkheia MCP server to your Claude Desktop or Cursor config. No infrastructure to run — served over stdio.

02

Tools available immediately

arkheia_verify and arkheia_audit_log appear as native tools in any MCP-compatible host — no further setup.

03

Call per-invocation

Your agent calls arkheia_verify after each model response. Detection happens at the tool layer, not in the API path.

Install promise

The public install path must stay boring: clear prerequisites, copy-paste client setup, visible key validation, and a first-run detection command. If a client needs special syntax, document it explicitly rather than making users debug stdio.

§ 03 · Enterprise Proxy API

Standalone HTTP API for enterprise pipelines

For pipelines that don't route through the cloud proxy, the Enterprise Proxy API exposes detection as a standalone FastAPI endpoint. Always returns HTTP 200 — never 4xx or 5xx — so detection failures never break your pipeline.

Request

POST /detect/verify
Content-Type: application/json

{
  "prompt":     "Explain the RFC 7230 spec",
  "response":   "RFC 7230 defines...",
  "model_id":   "gpt-4o",
  "session_id": "sess_abc123"  // optional
}

Response — always HTTP 200

{
  "risk_level":   "LOW",
  "confidence":   0.12,
  "detection_id": "det_abc123"
}

// On detection failure:
{
  "risk_level":   "UNKNOWN",
  "confidence":   0.0,
  "detection_id": null
}

Design contract

  • Always HTTP 200 — detection failures return risk_level: UNKNOWN, never 4xx/5xx
  • Pipeline-safe — detection never blocks or breaks downstream steps
  • Audit trail — every detection_id is persisted for later retrieval via arkheia_audit_log

§ 04 · Why Tool-Native

Works where endpoint redirection doesn't

Local models

Ollama and other local inference servers don't route through a cloud proxy. The MCP Trust Server calls detection after generation — no proxy needed.

Air-gapped environments

Enterprise deployments with strict network isolation can run the MCP Trust Server alongside the on-prem engine — detection never leaves the boundary.

Agent frameworks with fixed routing

Many agent frameworks have hardcoded API endpoints or SDK-level routing that can't be redirected. Tool-native detection bypasses this constraint entirely.

Complements the cloud proxy

The MCP Trust Server doesn't replace the cloud proxy — it extends coverage to integration paths the proxy can't reach. Both can coexist in the same deployment.

§ Get Started

One command. First verdict. Done.

Use a free Arkheia key or pass an existing one. The goal is a working detection receipt in minutes, not a platform evaluation.

One-command install

curl -fsSL https://arkheia.ai/install-mcp | bash

Or run directly via npx

npx @arkheia/mcp-server

Trust Layer Pricing

MCP Server and API Proxy use the same Arkheia trust-layer pricing. Choose the integration path that fits your workflow; the commercial model is the same.

Free

$0

1,500 detections/month

1 user · 7-day receipt history

Get started free →

Individual

$29/mo

More detections + 90-day receipts

Personal expense-friendly

Team

$299/mo

Shared projects, receipts, alerts

Invite teammates when it spreads

Business

$1k+/mo

Policies, approvals, exports, SSO

Turns visibility into governance

Get StartedView Full Platform