Product · dh gateway

Govern agents that don't use MCP.

Many agents run their own tool loop against an OpenAI-compatible endpoint. dh gateway sits between the agent and the model. The agent only changes its base_url, and every tool call the model asks for is decided and recorded.

Why it matters

The gap it closes

Agents built on the OpenAI SDK, LiteLLM, vLLM, Portkey or a framework's own loop have no tool server to wrap. The decision to call a tool happens in the model's response, and that's where the gateway sees it.

What it does

Capabilities

One line to adopt

Point the agent's OpenAI-compatible client at the sidecar. The model API key stays the agent's.

Any compatible upstream

OpenAI, LiteLLM, vLLM, Portkey or any /v1/chat/completions endpoint.

Tool calls decided in flight

The model's requested tool calls are checked against policy before the agent ever sees them as allowed.

Execution receipts

When the agent's runner signs a receipt under a key you pinned, the outcome is recorded as executed, not just reported.

Same record, same verifier

Records from the gateway verify with the same dh verify as the proxy's.

Honest about its view

The gateway sees what the model asked for. Without a runner receipt, an outcome is what the agent reported, and the record says so.

Get started

In practice

From the documentation.

dh gateway -c dh.toml --listen 127.0.0.1:8790 --upstream-url https://api.openai.com/v1

# in the agent
client = OpenAI(base_url="http://127.0.0.1:8790/v1", api_key=...)
Stated limits

What it doesn't do

  • Without runner receipts, the gateway can't confirm a tool actually ran; the record marks the outcome as reported.
  • The runner is trusted for execution facts when it signs receipts.

See it on one of your own workflows.

We'll put Decision Horizon in front of a real agent workflow of yours and replay the record with you.