indentapis / multiparty Goto Github PK

Multiparty provides fine-grained, programmatically defined permissions to control sensitive operations performed by AI Agents, without brittle if statements.

Sign up for the waitlist to give feedback and get access.

This repository serves as a central hub for the GitHub issues, use cases, and bug reports related to Multiparty for easier tracking and communication. For additional information, visit Multiparty.ai.

Run test suite:

go test ./pkg/engine.go

Run the simulator:

cd examples/simulator
npm i
npm run wasm
npm run dev

• What is it?
• Join the party
• Playground

We'd love to hear from you! Feel free to open an issue about a new use case or contact our team over email.

• Rules prescribe expectations for a system
  • Deterministic evaluation using Google's Common Expression Language (CEL)
• Actions run using WASM or Linux containers
• Plans specify ways to satisfy Rules using Actions

• Lightweight language wrappers around Rust/Go core
  • Python
  • JavaScript/TypeScript
  • Swift
  • Ruby
• Implements check and enforce commands
  • check quickly verifies assertions, e.g. don't execute if over budget
  • enforce provides async verification over longer windows, e.g. require user approval to be collected
• Support for local/embedded or remote engine
  • Embedded engine allows check without network connection
  • Remote engine provides centralized manage and control of rules

• Evalute Rules either locally or remote
  • Local for edge compute, testing, and single-tenant applications
  • Remote allows distributed and remote coordination
• Perform Actions in accordance to Plans
  • Dispatches WASM or containers for execution
  • Capture results and side effects to be used by Rules
• Durable storage for Plan and enforce calls
  • Execute Actions using either WASM or containers
• Signed decision chains skip the guesswork on why an Action was taken
  • Audit logs satisfy compliance requirements
• Budgets allow LLM-backed applications to deliver high quality UX with scalable costs

# langchain/libs/langchain/langchain/callbacks/human.py

from typing import Any, Callable, Dict, Optional
from uuid import UUID

from langchain.callbacks.base import BaseCallbackHandler
+ from multiparty.utils import async_input

def _default_approve(_input: str) -> bool:
    msg = (
        "Do you approve of the following input? "
-        "Anything except 'Y'/'Yes' (case-insensitive) will be treated as a no."
    )
    msg += "\n\n" + _input + "\n"
-   resp = input(msg)
-   return resp.lower() in ("yes", "y")
+   return async_input(msg)


def _default_true(_: Dict[str, Any]) -> bool:
    return True


class HumanRejectedException(Exception):
    """Exception to raise when a person manually review and rejects a value."""


class HumanApprovalCallbackHandler(BaseCallbackHandler):
    """Callback for manually validating values."""

    raise_error: bool = True

    def __init__(
        self,
        approve: Callable[[Any], bool] = _default_approve,
        should_check: Callable[[Dict[str, Any]], bool] = _default_true,
    ):
        self._approve = approve
        self._should_check = should_check

    def on_tool_start(
        self,
        serialized: Dict[str, Any],
        input_str: str,
        *,
        run_id: UUID,
        parent_run_id: Optional[UUID] = None,
        **kwargs: Any,
    ) -> Any:
        if self._should_check(serialized) and not self._approve(input_str):
            raise HumanRejectedException(
                f"Inputs {input_str} to tool {serialized} were rejected."
            )

When does Taylor's last meeting end today?

async function getMeetingTimes(range = 'today', view = ['time']) {
  await multiparty.enforce('get_calendar')

  const { meetings } = await calendar.getEvents(range)
  return meetings.map(event => {
    const { metadata, time } = await multiparty.enforce('get_calendar_event', {
      event,
      view
        })
    return { ...metadata, ...time }
  })
  return 
}

// Time
view.contains('time') && getFields(event, ['startTime', 'endTime'])

// Attendees
view.contains('attendees') && getFields(event, ['organizer', 'attendees'])

// Metadata
getFields(event, ['id', 'etag'])

// Check whether a resource name starts with a group name.
resource.name.startsWith("/groups/" + auth.claims.group)

// Determine whether the request is in the permitted time window.
request.time - resource.age < duration("24h")

// Check whether all resource names in a list match a given filter.
auth.claims.email_verified && resources.all(r, r.startsWith(auth.claims.email))