Case study library

Control patterns for real agent operations.

Use these cases to see how keys, permissions, approvals, budgets, and audit turn agent actions into governed operations.

Product proof

The pattern is consistent: control before action.

When agents can touch real services, MeterPlane decides who can act, what they can do, when approval is needed, how much they can spend, and what evidence remains.

10case analyses
5control layers
11public studies
Agent action POST /v1/services/{provider}/{operation}
  1. 01
    IdentityVerify the agent key, tenant, and target operation before provider access.
  2. 02
    AuthorizationEvaluate per-agent permissions, risk level, rate limits, and kill switches.
  3. 03
    ApprovalPause money-moving, publish, deploy, and destructive operations before credentials are read.
  4. 04
    Spend controlApply per-agent budgets, org limits, quotas, and enforced limits synchronously.
Credential access Only after layers 1-4 pass
05
Audit runs on every pathDenied, pending, and completed calls all write policy traces and usage evidence.
Credential control Raw provider keys stay vaulted

Database tools, AI APIs, and frontend apps all benefit from constrained gateway keys while provider keys remain vaulted.

Layer 1 + Layer 4
Approval control High-risk operations pause for review

Money movement, deploys, external publishing, and destructive database actions can pause for explicit approval.

Layer 2 + Layer 3
Spend control Budgets execute in real time

Per-agent caps, quotas, and organization limits define the spend boundary at request time.

Layer 4
Audit evidence Every decision leaves evidence

Policy decisions, MCP operations, data publishing, and usage events produce evidence your team can review.

Layer 5
00 5 layers

MeterPlane 5-layer control model

The product architecture study: identity, authorization, approval, spend control, and audit around every provider call.

ArchitectureCredential controlAudit
Read Full Case 01 9 sec

Database action approval

Destructive database operations can be denied or sent to approval before provider credentials are read.

IdentityAuthorizationApproval
Read Full Case 02 $82,314

Provider key budget control

Exposed AI provider keys become safer when raw credentials stay vaulted and budgets execute in real time.

CredentialRate limitBudget
Read Full Case 03 $440K+

Cross-agent transfer approval

Encoded instructions still resolve to operation types. token.transfer remains money movement and can require approval.

PermissionApprovalBudget
Read Full Case 04 $250K

Transfer spend limits

Parsing errors and large transfers are easier to contain when signing follows budget and approval controls.

SpendApprovalVault
Read Full Case 05 1,200+

Production DB policy control

Production database writes and deletes move from natural-language instruction to executable policy.

AuthorizationApprovalAudit
Read Full Case 06 CVSS 9.3

Read and publish separation

Sensitive context can be read while external publishing follows separate permission, approval, and audit.

Risk splitApprovalAudit
Read Full Case 07 4 mo

Per-agent budgets

Coding-agent loops stay manageable when every agent has a daily cap and the organization has a hard limit.

BudgetQuotaTrace
Read Full Case 08 $2.2K

Frontend gateway keys

Browser-side apps can expose constrained gateway keys while raw AI provider credentials stay vaulted.

CredentialRate limitBudget
Read Full Case 09 300 orgs

MCP operation audit

Built-in MCP, explicit email.send permission, and parameter audit make the send route inspectable.

MCPPermissionAudit
Read Full Case 10 65%

Governance and audit evidence

Security, spend, data, and compliance pressure all point to reconstructable action evidence.

All layersLoggingOversight
Read Full Case
Product takeaway

Each study turns public facts into a product control map: credentials, policy, spend, approval, and audit evidence.