Claude Desktop and MCP

Run the Mandate MCP server from npx or from the repo, and handle the guardian pause.

From npm

npx -y @yashjain99/mandate-mcp                 # stdio
npx -y @yashjain99/mandate-mcp --http 3111     # Streamable HTTP at http://localhost:3111/mcp

Environment: MANDATE_ACCOUNT; AGENT_PRIVATE_KEY (or the three CIRCLE_* variables) to execute, omit for read-only; optional GRAPH_API_KEY, BASE_SEPOLIA_RPC, ARC_TESTNET_RPC, MANDATE_STORE_DIR, MANDATE_ENV_FILE.

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "mandate": {
      "command": "npx",
      "args": ["-y", "@yashjain99/mandate-mcp"],
      "env": {
        "MANDATE_ACCOUNT": "0x…",
        "AGENT_PRIVATE_KEY": "0x…",
        "GRAPH_API_KEY": "…"
      }
    }
  }
}

Restart Claude Desktop. The server exposes eleven tools and ships its own instructions (the Mandate system prompt).

From the repo

{ "mcpServers": { "mandate": { "command": "pnpm", "args": ["--filter", "@yashjain99/mandate-mcp", "start"], "cwd": "/path/to/mandate" } } }

In-repo the server finds .env by walking up from its working directory and shares the workspace .data store with the web console, so a plan drafted in Claude Desktop shows up on the console's Activity page and its approval sheet can sign it.

The guardian pause

MCP is request and response, so the pause is explicit:

`prepare_step(planId, step)`

Agent-only steps execute right away. Guardian steps return approval: { text, guardian, chainId, deadline, nonce } and the model shows the text to you.

Sign the text

On the console's approval sheet with your Ledger (open the plan on the Activity page), or with any wallet that can personal_sign the exact text. In development, packages/core/scripts/dev-sign-approval.ts signs with GUARDIAN_PRIVATE_KEY and posts to the running console.

`submit_guardian_signature(planId, step, signature)`

The server verifies the signature against the cached text and the on-chain guardian. Only the signature is accepted from the client; text and guardian come from the server.

`execute_step(planId, step)`

Runs executeWithGuardian. Safe to retry: the chain is checked before anything is re-sent.

Pair with the skill

npx skills add yashj09/mandate -s mandate-agent

The mandate-agent skill tells the model the loop, the rules (never invent hashes, treat PerTxCapExceeded as intended behaviour, stop on awaitingGuardian) and the tool reference.

Smoke test without a model

cd apps/mcp && node scripts/smoke.mjs

Lists the tools over stdio. The handshake always succeeds; misconfiguration shows up as a tool error, not a crash.

On this page