AI Gateway

Guard every model call.Keep the evidence.

An OpenAI-compatible Gateway for requests and responses. Monitor sensitive data, tool misuse, and runaway context — then enforce when ready.

Gateway quickstart Limits & guardrails
Illustrative exchange · try a policy mode
One model-call boundary OpenAI compatible
Credential in contextFlagged
Disallowed tool callFlagged
Trace evidenceAttached

Watch matches. Keep traffic unchanged.

From first call to guarded traffic

Monitor first. Enforce on purpose.

  1. 01

    Route

    Use the Gateway as your client’s base URL.

  2. 02

    Monitor

    See policy matches. Keep calls unchanged.

  3. 03

    Enforce

    Redact, block, or abort supported matches.

  4. 04

    Observe

    Keep the same calls as diagnostic evidence.

What the Gateway controls

One boundary. Clear controls.

Sensitive data

Monitor, redact, or block supported data and credential patterns.

Tool-use policy

Check names and arguments. Detect repeated calls and error loops.

Context bounds

Cap history, tool definitions, input estimates, and completion tokens.

Usage limits

Bound requests, tokens, and cost by organization, project, or key.

Resilient routing

Provider routing, retries, fallback, and circuit breaking.

Trace evidence

Guardrail events stay attached for diagnostics and Findings.

Call-level guardrails here. Task-level diagnosis in Findings.

Start with one provider connection.

Route a client. Review the first policy matches.

Read Gateway docs