Apigee alternatives when the API caller is a model

You built an Apigee proxy, set a quota, and watched the analytics chart. The agent still created a refund by calling the payments host from a script. The proxy chart did not move. The customer was refunded twice.

This post shows what an API management suite records when the caller is a model, and what it misses when the call never enters a proxy.

What Apigee charges and what it can see

The Apigee pricing page meters API calls, environments, and proxy deployments. A Standard proxy is $20 per million calls up to 50 million, then $16, then $13. An Extensible proxy is $100, $80, and $64 on those same bands. A Base environment is $365 a month per region and can deploy Standard proxies only. Intermediate is $1,460 a month per region. The top environment tier on that page is $3,431 a month per region.

MCP in Apigee exposes your APIs as tools. Hosted analytics can split tools/list from tools/call and filter by MCP tool name. API insights uses those fields so you are not staring at one /mcp bucket. Those charts are built from traffic the message processors handled.

AI gateway capabilities also list extension processor policies for traffic that is not fronted by an Apigee proxy, including an agent's model and MCP tool calls. That feature exists. It is a separate setup from "we have a proxy, so we see everything."

Where the API suite still wins

Keep Apigee when you productize APIs for other developers. Quotas, shared flows, a developer portal, and per-tool latency belong there. If every agent must enter through your MCP proxy, the tool-name filter is a real control. API gateway alternatives covers that decision.

You also want Apigee when the environment price is justified by many proxies and a published SLA. A Base environment at $365 a month per region is a platform fee, not a per-agent seat.

What the suite does not record

A classic proxy records the request that arrived at the proxy. The model can still call a provider that you never proxied. No tool name appears, because there was no MCP call. The quota does not decrement.

Extension processors can attach policy to some of that traffic if you deploy them. Until you do, the analytics gap is the whole call. The fix for a chosen provider method is an allowlist next to the caller. Swytchcode checks the method you enabled and can block on arguments before the provider sees the body. Policy as code is the file. On the October 2026 pricing page, 1,000,000 live commands on Business are $149 a month. That does not include Apigee's portal, products, or environment SLA. It does include schema checks, credential injection outside the model, and a local audit via swy audit stats.

Logging every agent call is how you notice the side door. If the only log is the proxy, the side door looks like zero.

Walk one hour of traffic. A customer app calls your /orders proxy. Apigee counts it, and the tool-name filter stays empty because it was ordinary HTTP. An agent then calls the carrier API to buy a label, using a key in the worker. That request never had an Apigee proxy name. Your on-call graph looks healthy. The carrier invoice does not. Extension processors are how Apigee reaches some of this traffic, and they are worth it if you already pay for the platform and can deploy them. They are a poor reason to adopt Apigee only for one script. Count the side-door calls for a week before you buy another environment. If the count is small and the risk is a bad write, an allowlist is the smaller system. If the count is your entire partner API program, the proxy suite is the system you already meant to run.

See the call the proxy missed

Follow this step to install the CLI. Use one of these.

npm install -g swytchcode

irm https://cli.swytchcode.com/install.ps1 | iex

curl -fsSL https://cli.swytchcode.com/install.sh | sh

Confirm the install, then start the project.

swy --version
swy init

Search the registry for the provider the script calls directly. Add the read. Leave the refund method off the list until a person reviews it.

swy search <provider>
swy get <name from search>
swy list methods
swy add <canonical_id from swy list methods>

Connect the provider. There is no .env file in this project. Swytchcode keeps credentials in its own local store and reads them when a call runs.

swy auth connect <provider>
swy policy validate

Keep Apigee on the proxies you mean to productize. For the provider call that never enters one, require an enabled method and a policy before it is sent.

Related reading

• Log every API call an agent makes. Proxy analytics will not list a call that skipped the proxy.
• MCP gateways and execution layers. Tool discovery and the provider request are different hops.
• Credential brokering. A proxy quota does not keep the provider key out of the model.
• Human approval. A refund can wait even when the proxy would have allowed the HTTP call.

FAQ

Does Apigee bill an MCP tool call as an API call?

Yes, when that call is handled by an Apigee proxy. The pricing page charges API calls processed by proxies. Analytics can then split discovery from execution by tool name.

Can Apigee see agent traffic that is not behind a proxy?

Only if you use a feature that attaches to it. Google's AI capabilities page lists extension processor policies for traffic that is not fronted by a proxy, including model and MCP tool calls. A proxy you already deployed does not do that by itself.

Is $20 per million the whole Apigee bill?

No. That is the Standard proxy call price for the first 50 million calls. You also pay for the environment. Base is $365 a month per region. Extensible proxies cost more per million, and add-ons such as analytics are separate.

Should every agent call become an Apigee proxy?

Make a proxy when other developers or agents must discover a stable API you own. Skip the proxy when the agent is calling someone else's API and you need a check on that one method. Those are different deployments.