Technical overview

The architecture behind the decision room.

Manwe structures the work around a decision before asking AI to answer it: context first, roles second, disagreement third, record last.

Download beta Inspect records

Mac beta v0.6.5

System layers

Separate the jobs a normal chat interface blends together.

Mac app

A native beta workspace for starting runs, reviewing decision history, and exporting records.

Model layer

Runs can use local models or cloud models depending on the mode and configuration.

Evidence layer

User context, source material, and research notes are organized before advisors produce a verdict.

Advisor layer

Manwe separates the work into role-bound lenses instead of asking one assistant to do everything.

Dissent layer

The system pressures the emerging answer with objections, failure cases, and missing assumptions.

Record layer

Each serious run resolves into an artifact with verdict, evidence, risks, dissent, and next steps.

Pipeline

From question to record.

Frame Clarifies the choice, stakes, constraints, and what a useful answer must contain.

Gather Organizes the context and evidence the run should consider before answering.

Cast Builds a role-bound panel around the decision type and the user's goal.

Challenge Pressures the emerging answer with disagreement, risk checks, and alternative views.

Synthesize Turns the debate into a clear recommendation without hiding uncertainty.

Record Writes the result as a decision record that can be inspected, shared, or revisited.

Advisor pressure

Disagreement is part of the system.

Manwe does not treat the first plausible answer as finished. The system asks different roles to inspect the same decision from different angles, including failure cases, missing evidence, and practical next steps.

When users provide documents or notes, that context can become the primary material for the run instead of generic web-like reasoning.

Quick runs

For everyday choices that need a structured second mind.

Deeper runs

For strategy, product, career, money, or organizational decisions with real downside.

Source material

Users can add documents or notes when the answer should be grounded in their own context.

Exportable records

The output is designed to be read, shared, archived, and revisited later.