ClientOps Command

ClientOps Command — Evidence-Grounded AI Client Operations.

Comment

ClientOps Command — Evidence-Grounded AI Client Operations

Inspiration and problem

Client teams preparing for executive reviews often need to reconcile scattered account records, project delivery health, recent activity, and open commitments. A generic AI summary can sound persuasive while missing a business risk or citing evidence that does not exist.

ClientOps Command explores a more trustworthy approach: an AI agent that retrieves authoritative business data through a controlled tool interface, validates its evidence, and helps the user prepare for a client meeting.

What it does

ClientOps Command is a working, self-hosted Model Context Protocol (MCP) server and a companion Next.js demonstration application.

The MCP server implements MCP 2025-11-25 over Streamable HTTP and exposes five business READ/ANALYZE tools:

  • account.get_context
  • activity.get_recent
  • commitment.list_open
  • project.assess_delivery_health
  • meeting.prepare_brief

In Agentic Mode, Strands Agents TypeScript orchestrates authorized MCP tool calls using Amazon Bedrock. The application turns authoritative source records into an evidence-grounded executive briefing with risk findings, recommended next steps, expandable evidence, and a sanitized tool trace.

In Deterministic Mode, judges can reproduce a baseline briefing without an LLM or AWS credentials. Both modes require an explicit user action to execute.

How we built it

The system uses TypeScript, Node.js, Next.js, PostgreSQL, the official MCP server/client libraries, Amazon Bedrock, and Strands Agents TypeScript.

An evidence validator checks references in model-generated findings against source records retrieved during the actual tool execution. Deterministic domain rules, rather than the model, determine authoritative delivery-health status.

The agent is restricted to READ/ANALYZE MCP tools. A separate human-approval and trusted-execution architecture is included in the repository, but governed writes are disabled in the recommended public demo configuration.

Verified demonstration

In a recorded local browser demonstration using Amazon Bedrock with Claude Sonnet 4.6, a deliberate agent invocation returned HTTP 200 and produced:

  • Five successful business MCP tool calls
  • Six evidence-grounded findings
  • Six suggested next steps
  • Zero unsupported evidence references

A separate deterministic workflow produced a briefing with delivery-risk findings through two successful MCP calls.

The public source repository also passes standard GitHub Actions validation and a fresh-runner browser reproduction using a disposable PostgreSQL database and no AWS credentials.

Challenges and lessons

One important integration challenge was reconciling MCP tool names such as account.get_context with the stricter naming requirements of Bedrock Converse and the Strands tool registry. We introduced a reversible naming adapter so the model-facing names satisfy those constraints while MCP wire calls retain their original names.

We also worked through structured-output handling, strict TypeScript adapter types, bounded inference configuration, and separating agent suggestions from business authority.

These challenges reinforced a core lesson: reliable AI systems need explicit trust boundaries, verifiable evidence, and reproducible non-LLM tests.

Potential impact and next steps

The approach could help account managers, agencies, consultants, and service teams identify delivery risks earlier and prepare more reliable client reviews.

The current submission is a reproducible local demonstration with synthetic data. It is not a deployed Alexa+ Agent Skill, production Alexa+ integration, public SaaS service, or production-authenticated application.

Future work could add production authentication, additional governed business workflows, and real integrations while preserving the strict distinction between model-generated recommendations and human-authorized actions.

Source code and testing

Source: https://github.com/gabemart2020-collab/clientops-command-hackathon-2026

Demo video: https://youtu.be/HyUAR6tCCYQ

Judge instructions: https://github.com/gabemart2020-collab/clientops-command-hackathon-2026/blob/main/docs/DEMO.md

Development friction log: https://github.com/gabemart2020-collab/clientops-command-hackathon-2026/blob/main/docs/FRICTION_LOG.md

Built during the August–October 2026 hackathon submission period, with initial development beginning in September 2026.

Built With

Share this project:

Updates

Submission history