StrategyShifu

Inspiration

Complex cloud architecture and migration decisions are rarely solved by a single recommendation. Teams need to compare costs, technical constraints, operational trade-offs, and implementation risks before committing to a strategy.

We built StrategyShifu to explore a better model for human–agent collaboration: an AI agent that can work directly with a live decision-making application through WebMCP, while the user can see every change, review the reasoning, and make the final decision.

What it does

StrategyShifu is a WebMCP-powered decision workspace for cloud cost planning and data-platform migration strategy.

The current version supports two decision workflows:

• Databricks cost planning: Configure DWH, ETL, and development workloads and compare compute strategies using regional pricing, schedules, networking requirements, Databricks Genie compatibility, and project budgets.
• Redshift-to-Databricks migration planning: Compare Iceberg coexistence, Lakehouse Federation, UNLOAD to S3, and JDBC-based migration strategies using workload characteristics, downtime tolerance, synchronization plans, target formats, and execution readiness.

Through WebMCP, an agent can inspect the current project, modify inputs, test alternatives, calculate costs, compare strategies, and explain the recommendation. Every agent action is reflected immediately in the interface, allowing the user and the agent to work from the same decision context.

How we built it

StrategyShifu is built with Next.js, React, and TypeScript. WebMCP forms the interaction layer between the application and browser-based AI agents.

We followed several WebMCP best practices:

• Registered focused, task-oriented tools using document.modelContext.registerTool().
• Designed tools around user goals instead of exposing low-level UI actions.
• Used clear tool names, descriptions, and input schemas.
• Added runtime validation for agent-provided inputs.
• Returned both human-readable text and structured content.
• Included tool annotations to communicate read-only and mutating behavior.
• Used AbortSignal cleanup to prevent duplicate or stale tool registrations.
• Kept tool operations deterministic and safe.
• Used stable identifiers for selecting and updating workloads.
• Returned updated state after mutations so agents can continue without scraping the page.
• Provided clear errors and recovery guidance for invalid operations.
• Gracefully preserved the human interface when WebMCP was unavailable.

Most importantly, WebMCP tools and human controls operate on the same React state. There is no separate agent-only state or hidden decision model.

The application currently exposes:

• 17 WebMCP tools for workload configuration, pricing, cost estimation, strategy comparison, and recommendation retrieval.
• 7 WebMCP tools for migration inputs, capability evidence, and Redshift-to-Databricks assessment.

The recommendation engines remain deterministic and catalog-driven. The agent explores and configures scenarios, while the application owns calculations, validation, compatibility gates, scoring, and ranking.

Challenges we ran into

One of the main challenges was designing a WebMCP tool surface that represented meaningful decision operations rather than browser automation.

For example, instead of asking an agent to find and click several form fields, StrategyShifu provides tools such as add_workload, set_budget, compare_strategies, and get_recommendation. This makes agent interactions more reliable, understandable, and reusable.

Another challenge was synchronizing agent actions with human interactions. We solved this by connecting WebMCP tools directly to the same state update functions used by the visible interface.

We also had to balance agent flexibility with predictable recommendations. Agents can explore many scenarios, but technical compatibility, budget enforcement, pricing calculations, and migration scoring must remain reproducible and auditable.

Finally, migration decisions depend on evolving platform capabilities. We introduced structured capability evidence with documented-source links and a fallback mechanism so temporary documentation failures do not prevent an assessment.

Accomplishments that we're proud of

We are proud that StrategyShifu demonstrates WebMCP as a foundation for genuine collaboration rather than a collection of isolated agent commands.

Key accomplishments include:

• A shared live workspace where humans and agents operate on the same state.
• 24 task-oriented WebMCP tools across cost and migration planning.
• Structured tool responses that support multi-step agent workflows.
• Runtime input validation and actionable error messages.
• Safe tool registration with lifecycle cleanup.
• Deterministic and explainable recommendation engines.
• Full-project cost analysis across multiple workloads.
• Explicit Pass/Fail gates for budget and technical requirements.
• Transparent DBU and infrastructure cost breakdowns.
• Migration strategy scoring, alternatives, risks, and readiness evaluation.
• Suggested implementation steps and production cutover checks.
• Structured technical evidence supporting migration assessments.
• Share, copy, CSV, and Print/PDF export options.
• Responsive views for desktop and mobile users.

The agent does not simply summarize information. It can actively construct, modify, and evaluate a decision alongside the user.

What we learned

We learned that WebMCP works especially well for applications with structured state, repeatable workflows, and multiple possible decision paths.

Giving agents semantic tools is more dependable than asking them to infer application state from the visual interface. Structured inputs and outputs also make multi-step reasoning easier to validate and debug.

We also learned that agentic interaction should be separated from business logic. WebMCP provides the agent with controlled access to the application, while deterministic engines remain responsible for calculations and recommendations.

Finally, transparency is essential. Users are more likely to trust an agent-assisted recommendation when they can see the inputs, understand the failed requirements, compare alternatives, and reproduce the result themselves.

What's next for StrategyShifu

Our next goal is to turn StrategyShifu into a broader multi-platform strategy workspace.

Planned improvements include:

• Add more migration strategies across AWS, Azure, and Google Cloud.
• Support additional source and target platforms, including Snowflake, BigQuery, Synapse, PostgreSQL, and other data warehouses.
• Add cross-platform compute and architecture comparisons.
• Expand WebMCP tools into guided multi-step planning workflows.
• Add migration phase tracking and implementation checklists.
• Allow organizations to add their own policies, pricing agreements, and decision criteria.

Our long-term vision is for StrategyShifu to become a WebMCP-powered strategy layer across platforms—helping people and agents evaluate complex technology decisions together with transparency, consistency, and control.

Built With

• ai
• cloudflare
• next.js
• react
• typescript
• webmcp