TitanU Sovereign Agent Commander — Autonomous Mission Fabric (2026 upgrade)
Alexa+ track — self-hosted MCP 2025-11-25 Streamable HTTP interface plus an independent simulated voice/text assistant. No Amazon-certified skill or live Alexa device is claimed.
Why it matters
Most assistants return an answer. This system accepts an objective, delegates real bounded work to specialist agents operating over persistent user-owned task data, verifies material output, and produces a proof trail. It gives users control over both what agents may change and how they verify what happened. It is a public, deliberately limited demonstration of an architectural capability; proprietary TitanU orchestration code is not distributed.
What it actually does
Give the mission: "Prepare a TitanU product launch: find blockers, coordinate a plan, prepare outreach, and verify results." The system executes:
- Scope Scout selects relevant open tasks and hashes the input snapshot.
- Three concurrent agents: Risk Analyst flags blockers; Timeline Planner schedules by priority; Customer Comms prepares an unsent draft.
- Mission Control joins outputs and conditionally delegates remediation.
- Remediation Agent proposes up to three corrective tasks. It may create those tasks only if the user explicitly checks the local-write authorization box.
- Adaptive Planner automatically recalculates the schedule after authorized changes.
- Artifact Builder saves a real markdown mission brief, CSV schedule, and customer draft.
- Independent Auditor validates a SHA-256 event hash chain and checks every on-disk artifact against its stored digest.
A mission produces a persisted record of workers, activities, task mutations, integrity results and downloadable files. Local writes are off by default. No emails, external deployments, purchases or remote API actions occur.
Technical implementation
Python 3.10+ standard-library multithreaded workers with ThreadPoolExecutor; SQLite WAL for tasks, missions, events and receipts; asynchronous HTTP missions API; native HTML/CSS/JavaScript cinematic mission dashboard; optional browser voice capture with text fallback.
Actual MCP tool server: stateless JSON-RPC at POST /mcp over Streamable HTTP, negotiating protocol version 2025-11-25; supports initialize, notifications/initialized, tools/list and tools/call. Exposes titanu_mission_launch, titanu_mission_status, titanu_mission_verify, and titanu_task_list. Responses use JSON; optional server-to-client SSE is not implemented. This is an independently developed reference adapter, not a certification claim.
Working demonstration and testing
26 automated tests PASS, including HTTP lifecycle, MCP handshake and tool calls, parallel agent fan-out/fan-in, authorized task creation and adaptive replanning, permission-denied mode, duplicate prevention, path safety, and deliberately tampering with artifacts and event records to prove integrity failure is detected. A further headless browser interaction verifies eight agents, three downloadable artifacts, 25 audit events and zero JavaScript errors. The original four deterministic commands remain as additional examples.
Run without accounts or secrets:
git clone https://github.com/juliushill42/Titan-Hackathon-Moats.git
cd Titan-Hackathon-Moats/hackathons/amazon-2026-sovereign-commander
./titanu-amazon-commander-bootstrap.sh
# Browser: http://127.0.0.1:8787
./titanu-amazon-commander-bootstrap.sh proof
Originality, improvements and IP safeguards
Built and substantially upgraded during the 2026 hackathon window from the original simulated command workflows into genuine bounded, autonomous multi-agent coordination, evidence-driven replanning, persistent artifacts, an integrity-checking auditor and a Streamable HTTP MCP tool surface. These are new public clean-room components under MIT. No private EAISWARM source, private keys, internal endpoints, proprietary coordination algorithms or model weights are bundled.
Product feedback and limitations
Python stdlib and SQLite made local installation extremely simple and persistent output measurable. Browser speech recognition varies by device, so typed input is always offered. The default HTTP service binds to loopback, because the public demo has no remote-user authentication and should not be exposed to the internet. SHA-256 chains provide local tamper evidence, not globally signed/notarized history. Agent autonomy is bounded goal-directed orchestration, not general-purpose autonomous AI reasoning. Future work could add external identity, signed checkpoints, safe consent-based integrations and compatibility-tested Alexa+ experiences.
Public source
Built With
- agent-orchestration
- css3
- html5
- javascript
- json-rpc
- mcp
- python
- sha256
- sqlite
- web-speech-api