Skip to main content
Use this path when your runtime already supports MCP (Claude Desktop, Claude Code, Cursor, and similar clients) and you want real wallet operations exposed without writing a custom adapter first. If you have not run the SDK quickstart yet, do that first:

​
5-minute outcome

  1. Start the Cobo MCP stdio server
  2. Submit a pact and wait for owner activation notification
  3. Activate a pact from the MCP client itself
  4. Execute a blockchain action and track the resulting record
  5. Observe policy denial returned as tool content
  6. Retry with compliant params from suggestion
  7. Confirm via audit logs

​
Step 1: Install MCP extra

pip install "cobo-agentic-wallet[mcp]"

​
Step 2: Configure API access

export AGENT_WALLET_API_URL=https://api.agenticwallet.cobo.com
export AGENT_WALLET_API_KEY=your-api-key

​
Step 3: Start stdio MCP server

python -m cobo_agentic_wallet.mcp
To keep the runtime narrower, start with a subset: 
export AGENT_WALLET_INCLUDE_TOOLS=submit_pact,get_pact,contract_call,get_transaction_record_by_request_id,get_audit_logs
python -m cobo_agentic_wallet.mcp

​
Step 4: Register server in MCP client

Example (claude_desktop_config.json): 
{
  "mcpServers": {
    "cobo-agent-wallet": {
      "command": "python",
      "args": ["-m", "cobo_agentic_wallet.mcp"],
      "env": {
        "AGENT_WALLET_API_URL": "https://api.agenticwallet.cobo.com",
        "AGENT_WALLET_API_KEY": "your-api-key"
      }
    }
  }
}

​
Step 5: Run denial plus correction sequence

In your MCP client, execute this flow:
  1. list_wallets and get_wallet to identify the wallet the runtime should operate on
  2. submit_pact with an intent and PactSpec, then get_pact until the owner activates it in the Cobo Agentic Wallet app
  3. estimate_transfer_fee or estimate_contract_call_fee to preview the operation cost
  4. transfer_tokens, contract_call, message_sign, or payment with a stable request_id where applicable
  5. get_transaction_record_by_request_id or list_transaction_records to track the submitted operation
  6. Trigger a too-large transfer or disallowed contract call to observe the policy denial payload
  7. Parse suggestion from tool output and retry with compliant params
  8. get_audit_logs to verify allowed plus denied events
Policy denials are intentionally returned as normal tool content (not protocol-level tool failures), so the LLM can reason and retry. Non-policy API failures still surface as MCP errors.

​
Useful tool groups

  • Wallet discovery: list_wallets, get_wallet, list_wallet_addresses, get_balance
  • Pact lifecycle: submit_pact, get_pact, list_pacts
  • Blockchain execution: transfer_tokens, contract_call, message_sign, payment
  • Fee planning: estimate_transfer_fee, estimate_contract_call_fee
  • Tracking and recovery: list_transactions, list_transaction_records, get_transaction_record, get_transaction_record_by_request_id, list_recent_addresses
  • Audit and provisioning: get_audit_logs, create_delegation
Most MCP clients perform better when you expose a small, role-specific subset instead of the full surface.
  • Pact Drafting — submit_pact, get_pact, list_pacts
  • Execution — transfer_tokens, contract_call, estimate_transfer_fee, estimate_contract_call_fee, get_transaction_record_by_request_id
  • Observer — list_wallets, get_wallet, get_balance, list_transaction_records, get_audit_logs
If your runtime both drafts and executes, start with Pact Drafting plus Execution. Add Observer tools when you want the runtime to explain balances, inspect transaction outcomes, or review audit history on its own.

​
Go further

The MCP server exposes the same canonical toolkit surface as the other CAW integrations. You can still extend it with custom tools and combine it with direct SDK usage:
  • Add custom MCP tools — define additional MCP tools alongside Cobo’s for your own business logic (e.g. a check_price tool that fetches a token price before submitting a transfer).
  • Use the Python SDK for non-MCP operations — pact lifecycle management, policy configuration, and audit queries are all available via WalletAPIClient for operations you want to handle outside the LLM loop.
  • Use framework integrations for code-based runtimes — if you’re building a Python agent with LangChain, OpenAI Agents SDK, Agno, or CrewAI, use the framework-native toolkit instead of the MCP server. You get the same tools natively integrated with your runtime’s tool-calling mechanism.

Full MCP Integration Guide

Full 20-tool surface, config, and error semantics.

Python SDK

Direct programmatic access for operations outside the LLM loop.

LangChain

Framework-native toolkit for Python agents.