The governed path for every model call.
Agents keep their SDKs and change one base URL and one key. Every request is identified, checked, routed, forwarded with your provider key and recorded. Agents you can’t change are routed through your network, or through the gateway you already run.
Fraud Alert Triage · run 8f21c4
- Agent key
- ck_…7Qa2
- Model requested
- claude-sonnet-4.5
- Model used
- claude-sonnet-4.5
- Provider key
- Held by Clovie
- Tokens
- 3,812 in · 486 out
- Cost
- $0.0184
Every team calling providers directly means no one is in control.
Provider keys everywhere
Keys copied into code and CI give every agent the power of the whole account.
No single choke point
Policies, budgets and kill switches need one place where every request passes.
Gateways you already own
Platform teams have invested in Kong, Apigee or Envoy and won’t move every agent to a new address.
Clovie is both a gateway and a decision service. Use the Clovie gateway, or keep your own and let it ask Clovie before each request. Either way, one set of policies applies and coverage is labelled honestly.
Four ways in, one set of checks, every provider behind your own keys.
See it at work
Audit log
| Time | Agent | Model | Decision |
|---|---|---|---|
| 14:02:11 | Card Dispute Triage | gpt-4o | Deny |
| 14:01:58 | Fraud Alert Triage | claude-sonnet-4.5 | Allow |
- Policy
- Card numbers stay out of prompts
- Mode
- Enforcing
- Reason
- Card number found in the prompt
- Provider call
- Not made
Costs
| Agent | Team | Tokens | Cost |
|---|---|---|---|
| Card Dispute Triagethis request, blocked | Cards & Payments | 0 | $0.00 |
| Fraud Alert Triage | Cards & Payments | 18,420 | $0.31 |
| KYC Document Review | KYC Operations | 42,905 | $0.74 |
One change in the agent
Point any OpenAI- or Anthropic-compatible SDK at Clovie with the agent’s own key. Errors say which check stopped the request and why.
from openai import OpenAI
import os
client = OpenAI(
base_url="https://gateway.example.com/v1",
api_key=os.environ["CLOVIE_AGENT_KEY"],
)
reply = client.chat.completions.create(
model="anthropic/claude-sonnet-4.5",
messages=[{"role": "user", "content": "Summarise this dispute."}],
)HTTP/1.1 200 OK
{
"id": "chatcmpl-…",
"model": "anthropic/claude-sonnet-4.5",
"choices": [{ "message": { "role": "assistant", "content": "…" } }]
}What you get
Native APIs
Chat Completions, Messages and Responses, plus embeddings, images, audio, moderations, rerank, files and models, so OpenAI and Anthropic SDKs work unchanged.
Four ways to connect
Explicit endpoint, zero-code routing through your DNS or proxy, an agent front door for vendor agents, and Copilot Studio runtime protection.
Your gateway, our policies
Kong, Apigee, Azure API Management, Envoy and Istio, agentgateway, LiteLLM, Portkey, NVIDIA NeMo Relay and Bedrock AgentCore enforce through callouts; any OpenTelemetry source can be observed.
Routing and fallbacks
Route by tags, agent, team, environment or region, and fall back to other allowed models when one fails.
Access groups and rate limits
Named groups of approved models for agents, teams and keys, with request and token limits per organization, agent key and developer key.
Agent identity
One key per agent with automatic rotation, or workload sign-in with Kubernetes and cloud identities instead of stored keys.
Also in this layer
| Capability | What it does |
|---|---|
| MCP and A2A | Register MCP servers, govern tool calls, turn internal REST APIs into governed MCP tools, and route Agent2Agent calls through the gateway. |
| Coding agents | Start Claude Code, Codex, Gemini CLI or aider governed from the first request, with developer sign-in through your identity provider. |
| Built for the request path | Load shedding with Retry-After instead of slowdowns, fair sharing between organizations, and a liveness check that answers under load. |
Controls mapped to the frameworks you answer to
- SOC 2: SOC 2 Type II trust services criteria
- NIST AI RMF: NIST AI Risk Management Framework
- EU AI Act: EU Artificial Intelligence Act
- OWASP Top 10: OWASP Top 10 for LLM applications and agentic AI
Runs wherever Clovie runs
- SaaS
- Customer VPC
- On-premises
- Air-gapped
The hosted gateway serves SaaS. In self-hosted installs the gateway runs in your network; zero-code routing and gateways in your network are part of a self-hosted or hybrid install.
Questions
What happens if Clovie can’t check a request?
Your organization decides: continue with an alert (the default) or refuse requests until checks are back.
Do agents ever see our provider keys?
No. Agents use their own Clovie key. Only Clovie holds provider keys and uses them to call the provider.
Can we keep our existing AI gateway?
Yes. Supported gateways ask Clovie before forwarding each request and can send answers back through Clovie. Others can send telemetry to be observed and priced.
Which models can an agent use?
Only models your organization has enabled in the catalog and placed in the agent’s access group. Others return a clear error.
Put one governed path between your agents and every model.
See gateway and routing on your own agents, in the deployment you need.