Gemini Operations Navigator
daideguchi/gemini-operations-navigator
IA com guardrails de custo e aprovação humana para operações de suporte
Ver no GitHubServidor MCP que integra Gemini com controle de custos, logs de ações e checkpoints de aprovação humana para gerentes de operações de suporte. Bloqueia respostas ao cliente até revisão manual, exibindo ferramentas usadas, evidências citadas e projeção de gastos em tempo real.
Instalação
npx convez-ai daideguchi/gemini-operations-navigator1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Gerenciar filas de reembolsos e escalações com visibilidade de custos
- Rastrear evidências citadas por agentes de IA antes de respostas ao cliente
- Implementar checkpoints de aprovação em workflows de suporte automatizado
Veredito

Solução robusta e bem-documentada para operações sensíveis onde transparência de custos e aprovação humana são críticas. Implementação madura com traces compatíveis OpenInference/Arize e garantias de execução local sem sobrecustos.
README
A tradução para o português deste README ainda está na fila.
Gemini Operations Navigator
Gemini + MCP for support operations managers who need evidence-backed AI drafts, visible cost guardrails, and human approval before customer-facing action.
Live demo: https://daideguchi.github.io/gemini-operations-navigator/
YouTube demo: https://www.youtube.com/watch?v=kt34TmPsT4g
Submission package: SUBMISSION_PACKAGE.md
Architecture: ARCHITECTURE.md
Devpost field copy: rapid-agent/submission/devpost-submit-manual.md
Credit And Cost Boundary
Devpost sent an update on 2026-05-30 JST saying the hackathon-issued $100 Google Cloud credit codes for Google Cloud Rapid Agent Hackathon are no longer available because the pool was exhausted. The standard Google Cloud free trial remains a supported submission path.
This project must therefore stay on no-out-of-pocket routes only. Do not assume a $100 hackathon credit code will arrive, and do not use paid upgrades, prepay credits, auto-reload, or payment confirmation flows.
Judge Quick Read
Who is this for? Support operations managers handling refund and escalation queues.
What problem does it solve? AI support agents become risky when the manager cannot see which tools were used, what evidence was cited, how much the run may cost, and whether a refund promise is about to be sent.
How does it use AI? Gemini drafts from policy evidence, MCP tools retrieve and verify context, Arize Phoenix MCP exposes trace tooling, and the workflow blocks the customer reply until a human manager approves it.
What is verified? A terminal run, MCP tool trace, Arize/OpenInference-compatible span trace, cost ledger, approval checkpoint, Agent Builder-ready runtime manifest, actual Phoenix MCP stdio handshake, current Vertex rerun status, workflow UI, architecture diagram, narrated demo video, and public GitHub Pages review hub.
Demo


Demo video:
https://www.youtube.com/watch?v=kt34TmPsT4g
rapid-agent/media/gemini-operations-navigator-demo.mp4
Open in browser:
- https://daideguchi.github.io/gemini-operations-navigator/
- https://daideguchi.github.io/gemini-operations-navigator/rapid-agent/prototype/gemini-operations-navigator.html
- https://daideguchi.github.io/gemini-operations-navigator/rapid-agent/prototype/terminal-session.html
Run Locally
cd /path/to/gemini-operations-navigator
bash rapid-agent/scripts/run_google_local_checks.sh
Expected proof:
google_local_checks_ok
mcp_tool_calls=4
projected_cost_usd=0.021
cost_within_budget=True
human_approval_required=true
openinference_spans=4
partner_track=Arize
vertex_status=blocked_by_google_cloud_account_state
phoenix_mcp_tool_count=27
agent_builder_manifest=present
claim_boundary=local_workflow_verified_phoenix_mcp_verified_vertex_rerun_blocked_if_google_consumer_suspended_no_customer_send_claim
What It Shows
- MCP-style tool calls for policy, usage, draft, and customer reply.
- A grounded Gemini-style reply that cites evidence instead of promising a refund.
- A cost ledger that keeps projected spend visible.
- An Arize/OpenInference-compatible trace for every tool step, evidence ID, cost estimate, and approval boundary.
- A real Phoenix MCP server launch over stdio with
initializeandtools/listproof. - An Agent Builder-ready manifest for model, tools, MCP, budget, and approval policy.
- A current Vertex AI rerun report. If Google Cloud reports the project consumer is suspended, the project records that stopline instead of claiming a live generation.
- A high-risk customer-facing send blocked until manager approval.
Key Files
rapid-agent/case_data/- support ticket, policy excerpt, and tool catalog.rapid-agent/scripts/run_gemini_ops_agent.py- terminal workflow.rapid-agent/reports/mcp-tool-trace.jsonl- replayable tool trace.rapid-agent/reports/openinference-trace.jsonl- Arize/OpenInference-compatible span trace.rapid-agent/reports/cost-ledger.json- budget and projected cost.rapid-agent/reports/approval-checkpoint.md- human approval packet.rapid-agent/reports/vertex-gemini-live-proof.json- current Vertex AI Gemini rerun status.rapid-agent/reports/phoenix-mcp-runtime-proof.json- actual Phoenix MCP server handshake and tool list.rapid-agent/agent-builder/agent-builder-runtime-manifest.json- Agent Builder-ready runtime package.rapid-agent/prototype/terminal-session.html- terminal proof page.ARCHITECTURE.md- component and data-flow explanation.
Claim Boundary
Safe claim:
- A local Gemini/MCP workflow prototype runs against packaged case data and produces tool trace, cost ledger, approval checkpoint, workflow UI, Phoenix MCP runtime proof, Agent Builder-ready manifest, and natural English demo video.
- The current Google Cloud Vertex rerun status is recorded in
rapid-agent/reports/vertex-gemini-live-proof.json.
Not claimed yet:
- Production Google Cloud deployment.
- Final promotional-credit accounting.
- Unsupervised customer-facing action.