Gateway · the drop-in surface
One config line. Every LLM call governed.
Transparent LLM provider gateway for Anthropic, OpenAI, NVIDIA NIM, AWS Bedrock, and Google Vertex. Change one env var on the box where your agents run; every chat completion and tool call your agent makes through the gateway is evaluated against your Intended policies — no changes to agent code. For actions that never touch an LLM, govern them with the SDK by wrapping the action.
02 · The drop-in
Zero code changes — for the LLM traffic.
Set OPENAI_BASE_URL=https://gateway.intended.so/v1/openai (the per-provider path), along with INTENDED_KEY, INTENDED_TENANT_ID, and the X-Intended-Key header, on the box where your agents run. From then on, every chat completion and tool call your agent makes to that base URL flows through Intended — the code that talks to the model stays exactly the same.
The gateway governs what passes through the model. For actions that don't go through an LLM — a direct database write, a payment call, a deploy — there's nothing for the gateway to intercept, so you govern those with the SDK: wrap the action and submit the intent before it runs. Same policies, same audit chain.
03 · What it enforces
Policy, conformance, and fail-closed at the gateway.
Inline policy evaluation
Every tool call is evaluated against your declared policy set before the upstream LLM is invoked. Deny means the call never happens.
Sequence-aware
Recent activity is scored against your declared workflow. Out-of-order calls surface here instead of later.
Authority Tokens issued in-line
Approved calls receive a scoped, single-use Authority Token. Downstream connectors verify the token, not the gateway.
Every call sealed
Decision, rationale, token, and connector verification are hash-linked into the immutable audit chain.