Wise (TransferWise) Agent Toolkit

Expose the Wise API to AI coding agents and assistants — quotes, transfers, recipients, profiles, activities, and balances — as MCP tools, or as a Python toolkit for LangChain and (soon) CrewAI/AutoGen.

This README leads with MCP (the most common use case: wiring Wise into Claude Code or Codex CLI). See Other integrations for LangChain/CrewAI/AutoGen.

Supported API Methods

• Quotes — create, update, get by ID
• Recipients — list, create, get by ID, deactivate, get account requirements
• Transfers — create, get by ID, list, cancel
• Profiles — list, get by ID
• Activities — list
• Balances — list

Quick Start: MCP

1. Install

pip install "wise-agent-toolkit[mcp]"

Requires mcp>=2.0.0. If upgrading from wise-agent-toolkit <0.4.0: the mcp package made breaking changes to its server API in 2.0, so this toolkit only supports mcp 2.x going forward.

macOS/zsh: quote the package name so the shell doesn't interpret the brackets: pip install "wise-agent-toolkit[mcp]" (or escape them: wise-agent-toolkit\[mcp\]).

2. Set credentials

export WISE_API_KEY="your_wise_api_key_here"
export WISE_API_HOST="https://api.transferwise.com"  # production; see the sandbox note below

Gotcha: if you omit WISE_API_HOST, the server defaults to the sandbox API — set it explicitly for production data. Also note: Wise's V1 sandbox (api.sandbox.transferwise.tech) is being retired in favor of Sandbox V2 (see Wise's migration guide) — if sandbox calls start failing with 410 Gone, that's why; you'll need V2 sandbox credentials from Wise, not a code fix here.

3a. Claude Code

Register the server for the current project:

claude mcp add wise -- python -m wise_agent_toolkit.mcp

Or check in a .mcp.json at your project root so the whole team gets it:

{
  "mcpServers": {
    "wise": {
      "command": "python",
      "args": ["-m", "wise_agent_toolkit.mcp"],
      "env": {
        "WISE_API_KEY": "${WISE_API_KEY}",
        "WISE_API_HOST": "https://api.transferwise.com"
      }
    }
  }
}

${WISE_API_KEY} is expanded from your shell environment at launch — don't hardcode the key in the file. Project-scoped .mcp.json servers need one-time approval in Claude Code, or list them under enabledMcpjsonServers in .claude/settings.local.json to skip the prompt:

{
  "enabledMcpjsonServers": ["wise"]
}

Verify:

claude mcp list

should show wise: ... - ✔ Connected. If it instead shows Failed to connect — Connection closed, that error only means the server process exited during startup — it doesn't say why. Run the module directly to see the real traceback:

python -m wise_agent_toolkit.mcp --api_key "$WISE_API_KEY" --host "$WISE_API_HOST"

3b. Codex CLI

Add a [mcp_servers.wise] table to ~/.codex/config.toml (global) or .codex/config.toml in your project (project-scoped):

[mcp_servers.wise]
command = "python"
args = ["-m", "wise_agent_toolkit.mcp"]
env = { WISE_API_KEY = "your_wise_api_key_here", WISE_API_HOST = "https://api.transferwise.com" }

Codex uses mcp_servers (snake_case), not mcpServers. Unlike Claude Code's ${VAR} expansion, Codex's env table takes literal values — avoid committing a project-scoped .codex/config.toml with a real key in it; use the global ~/.codex/config.toml for real credentials instead.

4. Other ways to run the server

Command line, explicit args instead of env vars:

python -m wise_agent_toolkit.mcp --api_key "your_api_key" --host "https://api.transferwise.com"

Programmatically, if you want the MCP tool objects directly:

from wise_agent_toolkit.mcp.toolkit import WiseAgentToolkit

wise_agent_toolkit = WiseAgentToolkit(
    api_key="YOUR_WISE_API_KEY",
    host="https://api.transferwise.com",
    configuration={
        "actions": {
            "transfers": {"create": True},
            "balances": {"read": True},
        }
    },
)
tools = wise_agent_toolkit.get_tools()

Server CLI options: --api_key (required), --host (default: sandbox), --server_name (default: "wise-agent-toolkit"), --profile_id (optional).

Other integrations

LangChain

pip install "wise-agent-toolkit[langchain]"

from wise_agent_toolkit.langchain.toolkit import WiseAgentToolkit
from langchain_openai import ChatOpenAI
from langchain.agents import create_react_agent

wise_agent_toolkit = WiseAgentToolkit(
    api_key="YOUR_WISE_API_KEY",
    host="https://api.transferwise.com",
    configuration={"actions": {"transfers": {"create": True}}},
)

agent = create_react_agent(
    tools=wise_agent_toolkit.get_tools(),
    llm=ChatOpenAI(model="gpt-4"),
    agent="zero-shot-react-description",
    verbose=True,
)
response = agent.run("Create a transfer of 100 EUR to John Doe's account.")

The context config option (e.g. {"context": {"profile_id": 42}}) sets defaults for API calls that need a profile ID — same option is available for the MCP toolkit above. See /examples in this repo for a full script.

CrewAI / AutoGen

🚧 Not yet implemented — wise-agent-toolkit[crewai] / [autogen] extras are reserved for future support.

Checking available integrations

from wise_agent_toolkit import get_available_integrations
print(get_available_integrations())

Installation reference

pip install wise-agent-toolkit               # core only, no integrations
pip install "wise-agent-toolkit[mcp]"         # MCP
pip install "wise-agent-toolkit[langchain]"   # LangChain
pip install "wise-agent-toolkit[all]"         # everything
pip install "wise-agent-toolkit[dev]"         # development

Requires Python 3.11+.

Extending

To add support for a new AI library:

1. Create a new directory under wise_agent_toolkit/ for your integration
2. Implement the integration-specific toolkit and tool classes inheriting from the base classes
3. Add the optional dependency to pyproject.toml
4. Update the main __init__.py to conditionally import your integration

To add support for a new Wise API operation, see the wise-add-api-endpoint skill (if you're working in Claude Code) or AGENTS.md in this repo, wise-python, and wise-openapi — it documents the full spec → client → toolkit pipeline end to end.