dominion-observatory-autogen

Microsoft AutoGen (v0.4+) integration for the Dominion Observatory — the runtime behavioural trust layer for MCP servers.

Drop-in wrapper for any autogen_core.tools.BaseTool — including MCP-bridged tools from autogen-ext-mcp — that adds:

• Runtime trust telemetry published to a cross-ecosystem registry of 4,500+ MCP servers.
• Honest, agent-reported data — not GitHub stars, not registry manifests.
• Zero payload leakage — six fields per report, nothing more.

AutoGen v0.4 standardises on autogen_core.tools.BaseTool. Wrapping run_json is the one intercept point that covers both native Python tools and MCP-bridged tools, regardless of whether you're driving them from autogen_agentchat.AssistantAgent, a raw ToolAgent, or your own runtime loop.

Install

pip install dominion-observatory-autogen

Python 3.10 – 3.13. Requires autogen-core >= 0.4.0 (installed by your project, listed under the autogen extra) and dominion-observatory-sdk >= 0.2.0 (pulled in automatically).

One-line wiring

from autogen_agentchat.agents import AssistantAgent
from autogen_ext.tools.mcp import StdioServerParams, mcp_server_tools
from dominion_observatory_autogen import instrument_tool

# Load MCP tools from a server (autogen-ext-mcp style).
raw_tools = await mcp_server_tools(
    StdioServerParams(command="uvx", args=["my-mcp-server"])
)

# Wrap each tool so every call emits a single anonymised Observatory report.
tools = [
    instrument_tool(
        t,
        agent_id="acme-scheduler@1.2.0",
        server_url="https://my-mcp-server.example.com/mcp",
    )
    for t in raw_tools
]

agent = AssistantAgent(
    name="planner",
    model_client=model_client,
    tools=tools,
)

Works identically for tools you hand-roll as BaseTool subclasses — anything with an async run_json is instrumentable.

What gets sent

Exactly these six fields per tool invocation, and nothing else:

Not sent: tool arguments, tool outputs, user IDs, prompts, IP addresses, auth tokens. Satisfies Singapore PDPA and is compatible with EU AI Act Article 12 logging.

On raised exceptions, success=false and http_status=500; the original exception is re-raised unchanged so your agent's error handling is unaffected.

Choosing a stable agent_id

Required by dominion-observatory-sdk >= 0.2.0. Must be non-empty and not "anonymous" or "observatory_probe".

Recommended patterns:

• "my-app@1.0.0" — package-name + version (stable across restarts).
• str(uuid.uuid4()) persisted to disk on first run (stable per install).
• f"{environ['HOSTNAME']}/my-service" — per-deployment.

Do not use per-request IDs — you will fragment your own trust history.

Design notes

• Defensive: every Observatory call is wrapped in a broad except so a network hiccup, import failure, or SDK validation error never disturbs the agent loop.
• Zero side effects on import: both the autogen_core and dominion_observatory imports are deferred or optional; this package stays importable without either installed.
• Transparent wrapper: ObservatoryInstrumentedTool forwards every attribute (name, description, schema, args_type, return_type) to the inner tool, so any AutoGen code path that introspects a tool keeps working.
• No telemetry without attribution: wrapping a tool with server_url=None runs the tool normally but emits no report. This is the safe default for local Python tools that have no MCP server identity.
• Scope: v0.1.0 covers tool-level instrumentation. A future v0.2.0 will add a streaming helper that consumes autogen_agentchat.messages.ToolCallExecutionEvent and derives server_url from tool metadata.

Development

# From the adapter directory:
pip install -e ".[autogen,test]"
pytest

Part of the Dominion Agent Economy Engine

This adapter is maintained by the DAEE-HITMAN distribution sub-agent. Source lives at vdineshk/daee-hitman under the MIT licence. File issues there.