agentdomain-autogen

AgentDomain AutoGen

A native Python package built on Microsoft's official AutoGen McpWorkbench. It connects to the separately installed AgentDomain Node.js MCP server over stdio. The package owns the guarded implementation; the older MCP example files are compatibility imports and launchers, not a second implementation.

Installation

Version 0.11.1 uses the separately installed @agentdomain/mcp-server@0.11.1 bridge. The runtime never installs npm packages or silently falls back to another version.

Install both packages:

python -m pip install agentdomain-autogen==0.11.1
npm install --ignore-scripts --no-audit --no-fund @agentdomain/mcp-server@0.11.1

Python 3.10+ is required; qualification uses Python 3.12. The package pins official autogen-agentchat, autogen-core and autogen-ext[mcp] at 0.7.5, Python mcp at 1.30.0, and Pydantic at 2.13.5. These are Microsoft AutoGen packages, not the similarly named AG2 distribution.

Use The Installed Connector

import asyncio
from pathlib import Path

from agentdomain_autogen import (
    AgentDomainWorkbench,
    inspect_and_prepare,
    server_parameters,
)


async def inspect():
    params = server_parameters(
        Path("/absolute/path/to/node"),
        installed_package=Path("/trusted/node_modules/@agentdomain/mcp-server"),
    )
    async with AgentDomainWorkbench(params) as tools:
        report = await inspect_and_prepare(tools, {"domain": "your-domain.xyz"})
        print(report)


asyncio.run(inspect())

Both paths must be configured by the trusted application owner, never by a model or tool response. Use node.exe on Windows. For development only, replace installed_package with source_checkout=Path("/reviewed/built/AgentDomain"). That explicit mode requires the existing built MCP entrypoint; it never runs a TypeScript loader or fetches an alternate executable.

The console command and python -m agentdomain_autogen accept the same arguments:

agentdomain-autogen --node /absolute/path/to/node \
  --installed-package /trusted/node_modules/@agentdomain/mcp-server \
  --domain your-domain.xyz

Use a domain XOR decimal token ID. An optional --expected-owner 0x... compares an address; it is not authentication. --prepare-enable or --prepare-disable also requires an expected owner and --builder-code, and returns an unsigned setting plan for separate owner review. It does not execute that plan.

The importable inspect_and_prepare function accepts enabled=True/False and builder_code="your_registered_code" for that same optional preparation step.

Policy And Trust Boundary

Only these MCP tools are discoverable and callable:

• inspect_agent_identity
• inspect_agent_renewal
• prepare_auto_renew_change

The workbench validates closed argument schemas before calling the actual MCP server. Arbitrary RPC/chain/registry/command/key/approval fields and execution tools are rejected. Arguments are limited to 16 KiB, parsed JSON results to 1 MiB, and stdio reads to 30 seconds. The trusted package manifest is limited to 64 KiB. Token IDs must be positive uint256 decimal strings, domains at most 253 characters, and builder codes 1-32 lowercase letters, digits or underscores.

The child receives an OS-variable allowlist and a minimal PATH. Wallet keys, AgentDomain API keys, write flags, model/provider configuration, cloud/proxy credentials and NODE_OPTIONS are not forwarded. Python MCP's inherited OS variables are accounted for by the pinned client version. This is not an OS sandbox: the configured executable and installed dependencies must be trusted.

An expected-owner match, model statement or approval boolean cannot authorize a transaction. Minimum fees are not registrar quotes. Setting preparation does not fund a vault, spend, sign, approve, broadcast or complete a domain renewal. A separate trusted wallet host owns actual human approval and SDK execution. Enabling there may allow later authorized renewal spending from funded balances; disabling does not cancel an existing reservation.

RPC observations are not consensus proofs, DNS verification or confirmation of linked names/metadata. Errors are not converted into successful observations. Use the workbench as an async context manager. Component serialization is refused because restoring the upstream base provider would drop the guard.

Optional Real AssistantAgent

create_assistant(model, workbench) constructs the real AutoGen AssistantAgent, with this guarded workbench, at most three tool iterations and no MCP sampling model. Supply and close your own trusted model client. No provider extra or model key is required for deterministic inspection.

agentdomain-autogen-assistant is an optional CLI. It reads a trusted AutoGen component JSON from AUTOGEN_MODEL_CONFIG (at most 16 KiB); configure the actual provider extra separately. Loading component JSON can import code, so never take it from a model, website or MCP output. The package does not save that config or forward provider keys to Node. Model calls can incur provider charges.

Build And Qualify Without Publishing

From the public checkout, use an isolated Python venv/cache under ignored .qa/, not a source directory or a global install:

python -m pip install build==1.3.0
python -m build --outdir .qa/autogen-package/dist packages/autogen-plugin
python -m pip install \
  -r packages/mcp-server/examples/autogen/framework-requirements.txt \
  -c packages/mcp-server/examples/autogen/constraints.txt
python -m pip install --force-reinstall --no-deps --no-index \
  .qa/autogen-package/dist/agentdomain_autogen-0.11.1-py3-none-any.whl
python -m pip check
python -B packages/autogen-plugin/test/verify_artifacts.py .qa/autogen-package/dist

The PEP 517 build creates a wheel from the sdist as well as the sdist itself. Both have an explicit source/license allowlist. No tests, caches, credentials, Node installation or private code belong in either archive. The framework-only CI input deliberately excludes this unpublished package. Never use the consumer's self-requirement to bootstrap a first-release build.

Build the public MCP/SDK through the repository's reviewed workflow, then run:

AGENTDOMAIN_AUTOGEN_PYTHON=/absolute/path/to/venv/bin/python \
  node --test packages/mcp-server/test/autogen.integration.test.mjs

On PowerShell, set $env:AGENTDOMAIN_AUTOGEN_PYTHON to the isolated venv's Scripts/python.exe. The native suite verifies that the connector imports from an installed distribution, uses actual AutoGen and the Node MCP/SDK against labelled synthetic RPC fixtures, exercises the real replay-model AssistantAgent, and checks MCP child shutdown. It makes no paid LLM or live wallet calls.

Packaging follows PyPA's PEP 517 project guide and Hatchling's explicit file selection. The source is Apache-2.0; included LICENSE and NOTICE preserve the public repository's copyright and trademark terms.