VerifyAX MCP Server

The official Model Context Protocol (MCP) server for VerifyAX: a cloud-hosted bridge that gives your AI tools secure, real-time access to register agents, generate scenarios, run simulations, and read evaluations on the VerifyAX platform.

Getting started · Supported tools · Security · Deploy your own · Report an issue

---

The official VerifyAX MCP Server is a bridge between your MCP-compatible AI client and the VerifyAX agent-evaluation platform. Once configured, it lets you register agents, generate scenarios, run simulations, and read evaluations in natural language — without writing API scripts.

With the VerifyAX MCP Server, you can:

• Register and test agents (A2A or API) and confirm connectivity before evaluation.
• Generate scenarios from skill tags and wait for completion in one tool call.
• Run evaluations and read scores, transcripts, and credit usage without manual polling.

It complements (does not replace) the verifyax-api skill: the skill is for developers writing code; the MCP server is for conversational workflows.

Contents

• Supported clients
• Supported tools
• Before you start
• Install the MCP server
• How it works
• Example workflows
• Tips and tricks
• Data and security
• Troubleshooting
• Support and feedback
• Disclaimer
• For developers
• License

---

Supported clients

The VerifyAX MCP Server works with MCP-compatible clients that support Streamable HTTP or stdio:

Client	Setup reference
OpenAI ChatGPT	Connectors / MCP guide
Claude (Claude.ai, Desktop, and Code)	Claude MCP docs
Cursor	Cursor MCP docs
Visual Studio Code (GitHub Copilot)	VS Code MCP docs
GitHub Copilot CLI	About Copilot CLI
Google Gemini CLI	Gemini CLI MCP docs
Amazon Quick Suite	MCP integration guide

Any client that can connect via mcp-remote can also use the hosted endpoint at https://mcp.verifyax.com/mcp.

[!TIP] For step-by-step Claude setup and how this compares to the VerifyAX skill and SDK, see docs/using-verifyax-with-claude.md.

---

Supported tools

Twelve tools mapped to user intents. Blocking tools poll internally and return only when work completes (typically 30s–5min).

Area	Tools	Blocking
Agents	register_agent · list_agents · delete_agent	—
Scenarios	list_compatible_tags · generate_scenario · list_scenarios · delete_scenario	yes (generate_scenario)
Evaluation	evaluate_agent · list_recent_runs · get_run_details	yes (evaluate_agent)
Usage	get_usage_summary · preview_run_cost	—

[!NOTE] For tool descriptions (what Claude reads to pick a tool) and rationale, see docs/tool-descriptions.md. Package-level reference: packages/mcp-server/README.md.

---

Before you start

Requirements depend on how you connect.

Remote HTTP (hosted at mcp.verifyax.com)

• A VerifyAX API key (Settings → API Keys in the console)
• An MCP client with Streamable HTTP support, or Node.js 18+ to run the mcp-remote proxy

Local stdio (verifyax-mcp-server via npx)

• Node.js ≥ 20
• A VerifyAX API key in your client config (VERIFYAX_API_KEY)

The npm package exposes two binaries (verifyax-mcp-server and verifyax-mcp-server-http); pass the stdio binary name explicitly when using npx.

---

Install the MCP server

Remote HTTP (recommended)

Hosted endpoint:

https://mcp.verifyax.com/mcp

Send your VerifyAX API key on the initialize request. The hosted server does not store keys — each client session brings its own.

Cursor (native URL — v0.48+):

{
  "mcpServers": {
    "verifyax": {
      "url": "https://mcp.verifyax.com/mcp",
      "headers": {
        "Authorization": "Bearer sk-ver-api-..."
      }
    }
  }
}

Do not add a transport field — Cursor detects Streamable HTTP from the URL. You can also use "X-VerifyAX-API-Key": "sk-ver-api-..." instead of Authorization.

Claude Desktop / clients without native HTTP (mcp-remote):

Add to claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "verifyax": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcp.verifyax.com/mcp",
        "--transport",
        "http-only",
        "--header",
        "Authorization:${VERIFYAX_AUTH}"
      ],
      "env": {
        "VERIFYAX_AUTH": "Bearer sk-ver-api-..."
      }
    }
  }
}

Use --transport http-only because this server speaks Streamable HTTP only (no legacy SSE). You can also pass X-VerifyAX-API-Key: sk-ver-api-... via --header instead of Authorization.

Restart your MCP client after changing config.

Local stdio

Claude Code:

claude mcp add verifyax --env VERIFYAX_API_KEY=sk-ver-api-... -- npx -y -p @verifyax/mcp-server verifyax-mcp-server

Claude Desktop:

{
  "mcpServers": {
    "verifyax": {
      "command": "npx",
      "args": ["-y", "-p", "@verifyax/mcp-server", "verifyax-mcp-server"],
      "env": { "VERIFYAX_API_KEY": "sk-ver-api-..." }
    }
  }
}

Restart the client after changing config, then describe what you want in natural language.

---

How it works

Architecture and communication

1. Your MCP client connects to https://mcp.verifyax.com/mcp (remote HTTP) or spawns npx -p @verifyax/mcp-server verifyax-mcp-server (local stdio).
2. The client sends your VerifyAX API key on initialize (Authorization: Bearer … or X-VerifyAX-API-Key for HTTP; VERIFYAX_API_KEY env for stdio).
3. The server calls the VerifyAX API on your behalf. Blocking tools (generate_scenario, evaluate_agent) poll job status internally and return only when work finishes.

Permission and billing

• Actions are scoped to the workspace tied to your API key.
• Usage and credits are billed to your VerifyAX workspace, not to the MCP server operator.
• The hosted server sends no telemetry and does not persist API keys between sessions.

Self-hosting

Run the same HTTP server on your own infrastructure (e.g. Google Cloud Run). See deploy/gcp/README.md.

---

Example workflows

Once connected, describe tasks in natural language — the client picks the tool.

Agent setup

• Register: "Register my A2A agent at https://my-agent.example.com and confirm it's reachable."
• List: "What agents are registered in my workspace?"

Scenario authoring

• Discover tags: "List the skill tags I can use for an interview scenario."
• Generate: "Generate an info_exchange scenario tagged empathy and coordination."

Evaluation

• Run eval: "Evaluate agent X against scenario Y and summarize the scores."
• Review history: "Show details for my most recent simulation run."

Usage and cost

• Preview: "How many credits will it cost to run this scenario against my agent?"
• Summary: "What did my last 5 simulation runs cost?"

[!NOTE] Blocking tools can take 30s–5min. Do not call them repeatedly — wait for the result.

---

Tips and tricks

Add defaults to AGENTS.md

Reduce discovery calls and bad tag combinations by adding this to an AGENTS.md file in your project root (see the AGENTS.md convention for the format):

## VerifyAX MCP

When connected to verifyax:

- **MUST** call `list_compatible_tags` before `generate_scenario`
- **MUST NOT** combine QnA tags with other tags (QnA must be the sole tag)
- **MUST** use `preview_run_cost` when the user asks about credits before `evaluate_agent`
- Blocking tools (`generate_scenario`, `evaluate_agent`) take 30s–5min — do not retry manually

Use the skill for code workflows

For scripts, CI, or custom multi-step logic, use the verifyax-api skill or @verifyax/sdk instead of MCP tools.

---

Data and security

• All traffic to the hosted endpoint uses HTTPS (TLS).
• API key authentication; every action respects your VerifyAX workspace permissions.
• The hosted server does not store API keys — each session supplies its own.
• The MCP server sends no telemetry; it talks only to the VerifyAX API.
• Destructive tools (delete_agent, delete_scenario) permanently remove resources — confirm with the user before calling them.

API keys are managed in the VerifyAX console (Settings → API Keys). Revoke a key there to cut off access immediately.

---

Troubleshooting

• "VERIFYAX_API_KEY is not set" — the key isn't reaching the server. Check the env block (stdio) or Authorization / X-VerifyAX-API-Key header (HTTP).
• Authentication failed — the key is invalid, revoked, or from the wrong environment. Mint a fresh one in the console.
• Tool calls don't appear — MCP clients load tools at startup; restart the client (or start a new session) after adding the server.
• HTTP connection fails in Cursor — try mcp-remote with --transport http-only, or verify the URL ends with /mcp.
• Want logs? Set VERIFYAX_MCP_LOG_LEVEL=debug. Logs are structured JSON on stderr; stdout is reserved for the MCP protocol.

---

Support and feedback

This project is maintained by the VerifyAX team at Conscium.

• Bugs and feature requests: open an issue
• Development reference: CONTRIBUTING.md (maintainers and forks)
• External pull requests aren't accepted — issues are the best way to reach us

---

Disclaimer

MCP clients can register agents, generate scenarios, run evaluations, and delete resources using your VerifyAX API key and workspace credits. Use least privilege (scoped keys, minimal workspace access), review high-impact actions before confirming, and revoke keys you no longer need.

Model Context Protocol connects AI agents to external tools, which creates powerful workflows but also structural risks. Large language models can be vulnerable to prompt injection and related attacks. Only use trusted MCP clients and servers, and review which tools each agent can access.

---

For developers

This repository is a pnpm monorepo with two packages:

Package	Description
@verifyax/sdk	Typed TypeScript client for the VerifyAX REST API
@verifyax/mcp-server	MCP server (12 tools) built on the SDK

Requires Node ≥ 20 and pnpm 10.

pnpm install      # install workspace dependencies
pnpm build        # build all packages (topological order)
pnpm test         # run unit tests
pnpm lint         # lint
pnpm format       # format with prettier

Network-dependent suites: pnpm test:integration (live API, needs VERIFYAX_TEST_KEY) and pnpm test:conformance (spawns the built MCP server over stdio).

• MCP Inspector: docs/debugging-mcp-inspector.md
• Execution plan: docs/PLAN.md
• Architecture decisions: CLAUDE.md
• Cloud Run deploy: deploy/gcp/README.md

---

License

Apache-2.0. See LICENSE.