Everyone gave AI tools. Nobody gave it a supervisor.

That sentence kept me up at night. In 2026, assistants don't just answer questions - they act. They write, they delete, they send, they decide. And here's the question nobody building agents wants to ask out loud:

When your AI does something at 2 AM… who was watching?

We happily hand agents the keys to our documents, our logs, our inboxes - and then we just… trust the black box? I don't. So I built the thing the agent era is missing: a control plane.

What Console does

Console is a self-hosted MCP server (spec 2025-11-25, Streamable HTTP) that turns Alexa+ from an assistant into an operator - one that works under real governance:

Document Control - create, edit, and version documents by voice. Every edit is versioned. Nothing is silently overwritten.
Log Analysis - feed it application logs and ask "why did errors spike at 3 PM?" It ingests, clusters, detects anomalies, and writes the incident report.
Persistent Context - Console remembers. Across sessions, across conversations. "What did we decide about the migration last week?" — it knows.
Task Orchestration - describe a multi-step workflow out loud; Console plans the tool chain.
The Control Plane - this is the point of the whole project:
- Every tool call is audited - tool, args, result, latency, execution ID. Query it back by voice: "Console, what did you do today?"
- Sensitive actions hit a PermissionGate - they pause and wait for your explicit approval.
- A declarative PolicyEngine (policies.yaml) constrains what the agent can do at all.

How We built it

Python 3.11, official MCP SDK, Streamable HTTP transport - the Alexa+ open standard. Four capability modules sit behind a unified interceptor pipeline: PolicyEngine → PermissionGate → AuditLogger → tool. Persistence in SQLite so context and audit trails survive restarts. Log analysis uses real heuristics - sliding-window error detection and token-overlap clustering - no fake outputs, no mocked responses.

Everything runs locally with seeded demo data. No accounts, no keys, no excuses.

Challenges I faced

The MCP spec is brilliant and unfinished - which is exactly why this hackathon matters. Streamable HTTP session handling has sharp edges that stdio hides completely; I had to study reference implementations to get resumption right. Tool interception has no first-class middleware in the SDK, so I built a decorator-based interceptor to guarantee no tool bypasses the audit layer. And encoding "pending approval" as a tool result made me feel the spec's missing piece: a standardized deferred-execution result. My friction log documents all of it - because feedback is the real prize here.

What I learned

Giving an agent power is easy. Giving it power you can inspect, gate, and revoke is a discipline. We believe the next wave of consumer AI won't be won by the most capable agent - it'll be won by the most trustworthy one. Console is my bet on what that looks like: power, with a paper trail.

Built in ~13 hours for the Amazon Developer Hackathon, Alexa+ track.