Inspiration
What it does
How we buil## Inspiration
AI assistants are increasingly asked to help with decisions where a single answer can hide important trade-offs. Deliberation Council was built to give Alexa+ access to a structured reasoning service that considers a problem from multiple independent perspectives before producing a balanced synthesis.
What it does
Deliberation Council is a self-hosted Model Context Protocol (MCP) server that provides structured multi-perspective deliberation.
For each question, the council evaluates four perspectives:
- Practical — feasibility and real-world constraints
- Risk — possible failure modes and downsides
- Evidence — available support and reasoning
- Counterargument — challenges the emerging conclusion
The server then synthesizes these perspectives into a final result while preserving meaningful disagreement rather than hiding it.
How we built it
The project is implemented in Python as a self-hosted MCP server using Streamable HTTP and the MCP 2025-11-25 protocol specification.
It includes MCP initialization, tool discovery and tool calls, session handling, OAuth 2.1/PKCE-compatible authorization infrastructure, persistent SQLite state, bounded worker recovery, cancellation, restart-safe execution, and an audit trail.
The council engine runs the four perspectives independently and produces a weighted synthesis with explicit confidence and preserved dissent.
A deterministic provider is included for reproducible demonstrations, with optional local Ollama integration.
Challenges
One major challenge was making multi-perspective reasoning reliable rather than simply generating several similar answers. The council therefore keeps each perspective distinct and explicitly records disagreement.
Another challenge was building the service as a real MCP server rather than a visual prototype. The implementation includes protocol handling, persistence, authentication infrastructure, recovery behavior, and automated tests.
Accomplishments
- Working self-hosted MCP server
- Four-perspective deliberation council
- Explicit dissent preservation
- MCP 2025-11-25 support
- Streamable HTTP transport
- Persistent and restart-safe execution
- OAuth 2.1/PKCE-compatible security architecture
- Deterministic reproducible demo
- Optional local LLM support through Ollama
- 138 automated tests passing
What we learned
Structured disagreement can make AI reasoning more transparent. Instead of forcing every perspective into artificial agreement, preserving dissent gives the calling assistant more information about uncertainty, risk, and trade-offs.
What's next
Future work could expand the council with additional specialist perspectives, richer evidence sources, configurable council policies, and deeper integration with Alexa+ experiences.t it
Log in or sign up for Devpost to join the conversation.