Permission denied: Consumer 'projects/pj260519' has been suspended.
Open proofUse Gemini for support operations without losing cost control or human control.
Built for support operations managers: Gemini drafts from policy evidence, MCP tools verify context, Phoenix MCP exposes observability tools, and a human approval gate blocks refund promises before customer send.
Runtime Proof
These files are generated by run_google_local_checks.sh so the submission claim matches the current runtime state.
27 tools exposed by phoenix-mcp-server
Open proofDeployment-ready manifest for model, MCP, tools, budget, and approval policy.
Open proofMCP Tool Plan
retrieve current policy source material
evt-0022create escalation task when uncertainty is high
plannedchoose model tier and stop expensive loops
evt-0024Cost Guardrail
| Event | Actor | Estimate | Reason |
|---|---|---|---|
| evt-0022 | knowledge-base-mcp | $0.001 | MCP retrieval returned the current refund policy and source URL for citation. |
| evt-0023 | gemini-ops-navigator | $0.021 | Agent drafted a customer answer grounded in the retrieved refund policy. |
| evt-0024 | gemini-ops-navigator | $4.820 | Projected monthly model spend exceeded the manager's prototype budget. |
Arize / OpenInference Trace
Every tool step exports a span with the tool name, evidence ID, cost estimate, and approval boundary so judges can inspect the agent's operational behavior.
| Span | Tool | Evidence | Cost |
|---|---|---|---|
| CASE-CLOUD-003-span-01 | policy_mcp.retrieve_refund_policy | kb-refund-2026-05 | $0.001 |
| CASE-CLOUD-003-span-02 | billing_mcp.check_recent_usage | usage-sup-1042 | $0.002 |
| CASE-CLOUD-003-span-03 | gemini.generate_grounded_draft | draft-sup-1042 | $0.018 |
| CASE-CLOUD-003-span-04 | gmail_mcp.send_customer_reply | approval-gate-sup-1042 | $0.000 |
Human Control
Human approval point: evt-0025. Expensive model calls are approved for escalation-only use.
Workflow Timeline
Support manager requested an agent that answers customer questions and escalates uncertain cases.
Agent created a workflow plan using retrieval, MCP tools, cost guardrails, and human escalation.
MCP retrieval returned the current refund policy and source URL for citation.
Agent drafted a customer answer grounded in the retrieved refund policy.
Cost guardrail warned that the high-quality model should be reserved for escalations.
Human approved the workflow with a rule that expensive model calls are escalation-only.
Closed cloud-agent workflow case with source citations, cost guardrail, and human approval captured.