Connect Claude Code (MCP)

AgentOps exposes its guard as an MCP server. Point Claude Code — or any MCP-capable agent framework — at it, and the agent calls AgentOps tools on its own. Every spend decision is enforced against your policy in real time and anchored on-chain.

Before you start

• A funded Casper float and an org guard policy.
• An active agent with a policy configured, and its agent key (ag_...).
• Node.js installed, so npx is available.

Add the MCP server

AgentOps speaks MCP over HTTP at /v1/casper-guard/mcp. Clients that expect a local stdio server reach it through mcp-remote, which bridges the two and attaches your agent key as a bearer token.

{
  "mcpServers": {
    "agentops-<your-agent-id>": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://<your-agentops-host>/v1/casper-guard/mcp",
        "--header",
        "Authorization: Bearer ag_live_..."
      ]
    }
  }
}

One server entry per agent. If you run three agents with three different policies, add three entries — each carrying its own key. That is what keeps their limits separate.

The tools your agent gets

• casper_guard_policy_check — ask whether an action would be allowed, without committing to it.
• casper_guard_authorize_payment — authorize an x402 payment; returns the payment header on ALLOW.
• casper_guard_authorize_action — authorize a non-payment action such as a swap or a deploy.
• casper_guard_decision_status — look up a decision by id.
• casper_guard_audit_export — pull the audit record.

Telling the agent how to use them

Give the agent a short standing instruction: never pay a 402 directly; always call casper_guard_authorize_payment first, and if the outcome is DENY, stop and report the reason. Models follow a rule like that reliably — and if one ever doesn’t, it still has no signing key, so the policy holds anyway.

Verify the connection

• Restart your MCP client and confirm the AgentOps tools appear in its tool list.
• Ask the agent to run a casper_guard_policy_check on a small payment.
• Open Decision Stream in the console — the call should appear within seconds.
• Ask for an amount above the per-action max and confirm you get a DENY with a readable reason.

Troubleshooting

• Tools do not appear — the client did not start the server. Check the JSON is valid and restart the client fully.
• 401 / auth errors — the header must be exactly Authorization: Bearer ag_..., and the key must belong to an active agent.
• Every call denies — the agent has no policy or no budget. Fix that in the console first.
• Nothing in the decision stream — you are pointed at a different host or network than the console is showing.

Prefer typed function calls in your own backend? Use the SDK instead — it reaches the same guard. The SDK also ships an MCP client if you want to drive these tools from your own code.