Intent Gateway
Put your ArmorIQ policy on every tool call your agent's model returns, by changing one base URL. No SDK.
Intent Gateway sits between your agent and its model provider. Your agent's SDK sends its requests to gw.armoriq.ai instead of the provider. The gateway forwards each request with your provider key untouched, then checks every tool call in the model's reply against your agent's policy before your code receives it.
Your framework, prompts, tools and provider key stay as they are. The only change is the base URL.
What you get
- Every request on your dashboard. Requests, tool calls and token use appear in Observability under the agent that made them, whatever framework the agent runs on.
- Policy on every tool call. Name an agent in the URL and each tool call is checked against that agent's active policy. Allowed calls pass through unchanged. Blocked calls are removed from the reply, and the reply carries a note naming the rule that stopped them.
- A signed plan for every turn. Each model reply that asks for tools is recorded as a plan with a signed intent token, the same record the ArmorIQ SDK produces.
- Closed when in doubt. If ArmorIQ can't reach a decision for a named agent, that agent's tool call is withheld.
Two ways to connect
| Base URL | What happens |
|---|---|
https://gw.armoriq.ai/v1/t/<gateway-key> | Every request is recorded. Nothing is blocked. |
https://gw.armoriq.ai/v1/t/<gateway-key>/a/<agent-id> | Every request is recorded, and every tool call is checked against that agent's active policy. |
Start with the first to see your traffic, then add the agent to turn on enforcement.
Supported clients
| Client | What works through the gateway |
|---|---|
| OpenAI SDK | Chat Completions, streaming and non-streaming, including tool calls |
| Anthropic SDK | Messages, streaming and non-streaming, including tool use |
| Gemini | Gemini models through the OpenAI SDK, with your Gemini key |
Limits today
- Approvals. A hold rule works like a block. The gateway creates the approval request, but approving it does not release the agent's next attempt yet.
- Claude through the OpenAI SDK. Not supported. Use the Anthropic SDK for Claude models.
- Rate. A gateway key can make about 120 policy calls a minute. Each enforced model turn uses one plan, plus one check per tool call. Past that limit, a named agent's tool calls are withheld.
- Long requests. A request that runs longer than 5 minutes is cut off.
- Agents. Enforcement applies to an agent that already exists in ArmorIQ and is active. See Add an agent.
Next steps
- Quickstart - Connect an agent in a few minutes.
- How enforcement works - What the gateway does with each tool call, and what happens when ArmorIQ can't decide.