Common Scenarios

Real problems Axemere solves

Not a feature list. These are the specific situations where a gateway layer pays for itself, with the console steps or policy config that solves each one.

Budget control

Stop runaway AI spend before it happens

The problem

A single misconfigured script or a prompt loop can blow through thousands of dollars of API budget in hours. By the time you get an invoice, the damage is done.

How Axemere solves it

Set a monthly budget limit per workload. The gateway enforces it in real time, blocking requests the moment the threshold is hit, not at billing time. Configure a Slack or email alert at 80% so you know before you're over.

Console
Policies → New Rule → Condition: workload_spend_usd_mtd ≥ 500
Effect: Deny · Reason: "Monthly budget cap reached"
Alert Channels → Add alert at 80% threshold
Policy DSL reference
Model routing

Route requests to the right model for the job

The problem

GPT-4o is overkill for summarization tasks. Haiku is too slow for latency-sensitive code completion. Hardcoding model choices in application code makes changing them a deployment event.

How Axemere solves it

Write routing rules in YAML at the gateway. Route by workload ID, request label, or any attribute of the request context. The gateway selects the connector and credential, and can rewrite the target path and model. No application code changes required: just push a new policy bundle.

YAML (Free Gateway / CP-connected)
# Route 'wl-summarizer' to Anthropic/Haiku; everything else uses org default targets: - id: route.summarizer.haiku priority: 100 when: field: context.workload_id equals: wl-summarizer effect: decision: allow select_connector: connector_id: anthropic select_credential: credential_id: cred-anthropic mutate: action.target_path: /v1/messages action.params.model: claude-haiku-4-5 # Note: action.params.model via mutate applies to explicit API mode (POST /v1/gateway/actions). # In transparent proxy mode, achieve model selection by routing to a credential/connector # that is pre-configured for the desired model — do not rely on mutate to override it.
Policy DSL reference
Credential management

Rotate AI provider keys without touching application code

The problem

When an AI provider key is compromised or expires, every service that has it hardcoded needs an emergency update. In a distributed system, that's a fire drill.

How Axemere solves it

Applications authenticate to Axemere with a gateway API key. Provider keys live only in the gateway. Rotate a provider key once at the gateway and all workloads immediately use the new one: zero code changes, zero downtime.

Console
Credentials → Select credential → Edit → Update API key → Save
All workloads pick up the new key on the next request.
Credentials guide
Audit trail

Prove what happened on every AI request

The problem

An enterprise customer asks: 'Can you show us every AI request your system made last quarter, what model it used, what it cost, and that every request went through an authorized, governed pathway?' Today, the answer is 'no'.

How Axemere solves it

Every request through the gateway generates a signed execution record: provider, model, token counts, cost, policy decision, workload, project, and attribution labels. Records are queryable, exportable, and (with the Ledger package) cryptographically verified via Merkle proofs.

Console
Records → Filter by project, workload, date range → Export CSV or XLSX
Analytics → Attribution → drill into provider, model, workload breakdowns
Records guide
Team access

Give every team their own AI budget and policy scope

The problem

The ML platform team and the product team share one OpenAI account. There's no way to see who spent what, enforce team-level limits, or block one team's experiment from affecting another's production workloads.

How Axemere solves it

Create a workload per team or per application. Each workload gets its own budget limit, policy scope, and attribution labels. The central operations team sees the fleet view; each team sees only its own data. Rotate credentials per team without touching other workloads.

Workloads guide
Approval workflows

Put a human in the loop before a sensitive request executes

The problem

Some AI actions (sending a customer-facing message, modifying a financial record, calling an external API) shouldn't execute without a human review. But adding approval logic to every application is expensive and inconsistent.

How Axemere solves it

Write a policy rule with require_approval for matching requests. When triggered, the gateway returns HTTP 202 immediately with an approval ID; the original request is held pending, not executed. An operator reviews it in the console (Govern → Approvals) or via API, approves or denies with a reason. Once approved, future requests matching the same workload and action type are pre-authorized and execute normally. No application logic required beyond handling the 202 and retrying on approval.

Console
Govern → Approvals → filter by Pending
Select a request → Approve (no body required) or Deny (reason required)
Approved requests pre-authorize future requests for the same workload + action type.
Approval workflows guide

Try it yourself in 5 minutes

The free gateway runs on macOS, Linux, Windows, and Docker. No account required. Start routing AI traffic locally and apply your first policy in minutes.