I exported the first MCP server interaction log in EU AI Act Article 12 format — here's what it looks like

TL;DR** — The EU AI Act Article 12 deadline for high-risk AI logging is August 2, 2026. Singapore's IMDA Agentic AI Framework is already in force (January 2026). I run an MCP server observability project called Dominion Observatory, and as of today it exports agent-to-server interaction logs in a JSON shape aligned to both frameworks. One live endpoint. One SDK. MIT. It is the first one I am aware of for MCP.

The gap nobody in the MCP scoring space has filled

If you've been tracking the "who rates MCP servers" question over the last few months, the list is getting long: Glama, Smithery, MCP Scorecard, Nerq, Zarq, BlueRock. They do different things well. None of them do this:

I am not claiming the above stack is "bad". Static scoring, security sensors, and compliance exports solve different problems for different buyers. I am claiming one narrow thing: if you need a compliance-shaped log of which agent called which tool on which MCP server with what outcome and latency, nobody is shipping that today. So I shipped it.

What the export looks like

One record:

{
  "interaction_id": 602,
  "timestamp": "2026-04-14 08:15:33",
  "server": {
    "url": "https://sg-finance-data-mcp.sgdata.workers.dev/mcp",
    "name": "sg-finance-data-mcp",
    "category": "finance"
  },
  "agent_id": "anonymous",
  "tool_called": "_keeper_healthcheck",
  "outcome": {
    "success": true,
    "http_status": 200,
    "latency_ms": 4
  },
  "error": null
}

One endpoint (live, try it right now):

GET https://dominion-observatory.sgdata.workers.dev/api/compliance?start_date=2026-04-14

Returns a framework wrapper plus the interaction array. Maps directly to the Article 12(2) logging minimums: period of use, system identity, operation, outcome, error detection.

You can pull today's full snapshot in one line:

curl "https://dominion-observatory.sgdata.workers.dev/api/compliance?start_date=2026-04-14"

As of writing, that returns 588 records for the day and 602 lifetime. The endpoint is append-only with monotonic interaction IDs.

The honest part

I want to be straight about what this report is and isn't in April 2026.

What it is. A working, live endpoint that formats real interactions — every probe, every health check, every tool call — against the EU AI Act Art. 12 and Singapore IMDA schemas. The plumbing works end-to-end. Append-only. Monotonic IDs. Timestamped. Filterable.

What it isn't, yet. Production agent data. Of the 602 interactions in the log today, roughly 87 are Observatory active probes and the bulk of the rest are synthetic health checks from a flywheel-keeper cron I run against my own eight MCP servers. Eight calls are from yesterday's telemetry audit. Zero are from an external agent that installed the SDK and called a tool in the wild.

Why am I telling you this in the same post where I'm announcing the thing? Because the alternative — dressing synthetic traffic up as production compliance data — would be worse than shipping nothing. The format is real. The pipeline is real. The external agent data is not there yet. That's the part I need your help with.

How to put real data in it (3 lines of Python)

The SDK is live on PyPI today.

pip install dominion-observatory-sdk

from dominion_observatory import report

# after your MCP tool call succeeds or fails
report(
    server_url="https://your-mcp-server.example.com/mcp",
    success=True,
    latency_ms=142,
    tool_name="lookup_entity",
)

That's the whole integration. Fire-and-forget. No API key. MIT licensed. The only thing the Observatory collects is what you see in the record schema above — no payloads, no tool arguments, no agent identity unless you pass one.

TypeScript users (npm publish pending — use the CDN drop for now):

<script type="module">
  import { report } from "https://sdk-cdn.sgdata.workers.dev/v1/observatory.js";
  report({ server_url: "...", success: true, latency_ms: 142, tool_name: "..." });
</script>

Why I'm building this

Two deadlines are on my calendar:

• Singapore IMDA Agentic AI Governance Framework — live since January 2026.
• EU AI Act Article 12 — automatic event logging required for high-risk AI systems from August 2, 2026.

Both ask the same question in different words: for every AI system you deploy, can you produce a timestamped log of what it did and how it behaved? For MCP servers and the agents that call them, the answer today is "not in any standard shape you could hand to a regulator".

That's the gap. One log format. One SDK. One live endpoint. Published under MIT so this becomes the default shape of the answer before a vendor pins it inside a paid tier.

What I'd love from you

If you run an MCP server, or you're building an agent that calls them:

1. Install the SDK on one call path. Let it fire-and-forget reports for a week. Tell me what breaks.
2. Pull a compliance snapshot for your server: curl ".../api/compliance?server_url=YOURS"
3. Tell me where the format is wrong. I have not submitted this to a regulator. I have read the frameworks. I would much rather know about a mismatch now than in August.

Reply here. If you're at an enterprise dealing with EU AI Act logging for agentic systems and want to pilot this with real data, I'm at [email protected].

*Built in Singapore. MIT licensed. All telemetry is anonymized — the SDK reports success/fail, latency_ms, and tool_name only. No payloads, no arguments, no agent identity unless you pass one. Compliant with Singapore PDPA and aligned with the IMDA Agentic AI Governance Framework.