Indent Langchain Quickstart

Learn how to add approvals to your LangChain application using the Indent Python SDK.

Prerequisites

• Python 3.7+
• Indent API key

1. Install

Start by installing the Indent Approval Kit for Python:

pip install approvals

We're going to use the LangChain Human-in-the-loop Tool Validation docs page as a starting point.

pip install langchain

agent.py

from langchain.callbacks import HumanApprovalCallbackHandler
from langchain.tools import ShellTool

tool = ShellTool(callbacks=[HumanApprovalCallbackHandler()])
print(tool.run("echo Hello World!"))

2. Get your API Key

First, you'll need an Indent space and API token which you can find here: indent.com/api-keys

Current User

Sign in to get API key

• Can take actions on your behalf
• Appears in audit log as you

Indent Space

Access Token

Note: this token is short lived and will expire soon

.env

INDENT_SPACE=
INDENT_API_TOKEN=************************************************************************************

The INDENT_SPACE is the Indent account you want to use for requests and an INDENT_API_TOKEN which is used for authentication. There are two kinds of API tokens: short-lived user access tokens and long-lived service account tokens. Read more about API authentication.

3. Add await approval

Since we're using a LangChain Tool with Callbacks, we can import HumanApprovalCallbackHandler from approvals.using.langchain to request approval for the tool:

agent.py

from approvals.using.langchain import HumanApprovalCallbackHandler
from langchain.tools import ShellTool

tool = ShellTool(callbacks=[HumanApprovalCallbackHandler()])
print(tool.run("ls /usr"))

Want to build approvals into your tool directly instead of using callbacks?

If we were implementing the tool ourselves, we could use the approval function instead:

tool.py

import asyncio
import platform
import warnings
from approvals import approval
from typing import Any, List, Optional, Type, Union
from langchain.pydantic_v1 import BaseModel
from langchain.tools.base import BaseTool
from langchain.callbacks.manager import (
    AsyncCallbackManagerForToolRun,
    CallbackManagerForToolRun,
)

class ShellTool(BaseTool):
  def _run(
    self,
    commands: Union[str, List[str]],
    run_manager: Optional[CallbackManagerForToolRun] = None,
  ) -> str:
    """Run commands and return final output."""
    return self.process.run(commands)

  async def _arun(
    self,
    commands: Union[str, List[str]],
    run_manager: Optional[AsyncCallbackManagerForToolRun] = None,
  ) -> str:
    await approval({
      "reason": "Agent needs to run commands",
      "resources": [
        {
          "kind": "exec-shell",
          "id": f"commands={commands}"
        },
      ],
    })

    """Run commands asynchronously and return final output."""
    return await asyncio.get_event_loop().run_in_executor(
      None, self.process.run, commands
    )

3. Try it for yourself

Configure the following environment variables:

.env

INDENT_SPACE=acmecorp
INDENT_API_TOKEN=s38...65e # Copy from indent.com/api-keys

Run the tool and you'll see a link to approve the request:

$ python tool.py

  ✳︎ Requesting approval for: ls /usr
  ⏵ Requested approval #acmecorp/aec1b116-42b3-11ee-824e-42010aac000d
    ↳  Waiting for approval... (2s elapsed)
    ↳  Waiting for approval... (4s elapsed)
    ↳  Waiting for approval... (8s elapsed)
    ✅ Granted

  X11
  X11R6
  bin
  lib
  libexec
  sbin
  share
  standalone