Skip to main content

​
Overview

The OpenAI integration has two practical paths:
  1. Python adapter — use the built-in CAW OpenAI adapter
  2. TypeScript custom tools — wrap the CAW TypeScript SDK in OpenAI Agents SDK tool definitions
The Python integration exposes three levels of abstraction:
  1. build_function_tools() — raw function tools
  2. get_cobo_tools() — bring-your-own agent
  3. create_cobo_agent() — preconfigured agent with denial-aware guidance

​
Quick Start

1

Create the agent

from agents import Runner
from cobo_agentic_wallet.client import WalletAPIClient
from cobo_agentic_wallet.integrations.openai import (
    create_cobo_agent,
    create_cobo_agent_context,
)

client = WalletAPIClient(
    base_url="https://api.agenticwallet.cobo.com",
    api_key="<api_key>",
)
agent = create_cobo_agent(client=client, model="gpt-4.1-mini")
context = create_cobo_agent_context()
2

Run the agent

result = Runner.run_sync(agent, "List my wallets", context=context, max_turns=8)
print(result.final_output)

​
Tool Selection

To restrict the CAW tools exposed to the agent, pass include_tools or exclude_tools: 
agent = create_cobo_agent(
    client=client,
    model="gpt-4.1-mini",
    include_tools=["submit_pact", "get_pact", "contract_call", "get_transaction_record_by_request_id"],
)
If you need the raw filtered tool list instead of a prebuilt agent, call get_cobo_tools(client, include_tools=[...]).

​
TypeScript Path

When using the OpenAI Agents SDK in TypeScript, use @cobo/agentic-wallet as the backend and expose CAW operations through a small set of custom tools. Keep the same role-based split used elsewhere in the docs:
  • Pact Drafting: submitPact, getPact
  • Execution: transferTokens, contractCall, fee estimation
  • Observer: transaction record lookup and audit logs
The CAW repo does not currently ship a dedicated TypeScript OpenAI adapter layer equivalent to create_cobo_agent(). The recommended pattern is to keep CAW inside your own tool functions so you control tool naming, retry logic, and the exact surface exposed to the model.

​
Three Levels of Integration

Function Tools

build_function_tools() returns raw function tools for maximum flexibility.

Tool Set

get_cobo_tools() returns tools ready for any OpenAI agent.

Full Agent

create_cobo_agent() returns a preconfigured agent with denial-aware system prompt.

​
Denial Behavior

Policy denials are preserved as structured guidance and fed back into subsequent agent instructions for retry decisions.
When a tool returns a policy denial, the OpenAI agent receives the structured denial (code, reason, details, suggestion) in its context. The agent’s system prompt instructs it to parse this feedback and retry with compliant parameters.

Source Code

src/cobo_agentic_wallet/integrations/openai/agent.py