Govern people and agents in a single layer: ACID financial pre-authorization, multi-LLM routing and tamper-proof audit — between your applications and any provider (OpenAI, Anthropic, Google or self-hosted).
# gw_… key — models, rate limit and IP already on the token curl https://api.ai-gateway-core.com/v1/chat/completions \ -H "Authorization: Bearer gw_live_••••" \ -d '{ "model": "anthropic/claude-opus", "messages": [{ "role":"user", "content":"summarize the contract" }] }' # 1 reserve → 2 provider → 3 settle
Multi-LLM
providers + self-hosted
ACID
reserve → settle on every call
SHA-256
chained, verifiable audit
0
keys exposed to the caller
// running AI without governance is expensive
Every team spins up its own provider key, with no limit and no visibility. The bill only shows up at the end of the month.
Reserve → settle cycle on every call. With no budget left in the area, the call does not happen (HTTP 402).
Every automation — n8n flow, agent, script — gets its own key and spends without limits. Nobody knows which agent did what, or how to cut one off without taking down the rest.
Each agent is a headless identity with an owner, environment, models, IPs and its own token. Budget and rate limit per agent; suspend or rotate in isolation.
With no reliable record of what was spent and why, audit and compliance turn into fragile manual work.
Append-only trail chained by SHA-256, verifiable, with PII redacted before any persistence.
// what the gateway does on every call
Managed providers (OpenAI, Anthropic, Google) and self-hosted models on your own infrastructure — vLLM, Ollama or internal LLMs — behind a single compatible API. Switch providers without changing code; credentials stay encrypted server-side (AES-256) and are never revealed to the caller.
consumed by
routed to
governance on every call · keys encrypted server-side
active agents · Finance area
owner: maria@headless identity · own gw_ token · auditable by owner
Headless automations — n8n flows, agents and scripts — are first-class identities, separated from people. Each agent has a responsible owner, purpose, environment, allowed models, IPs and its own rotating gw_… token — with independent budget and rate limit.
Not a thin proxy. Before calling the AI, the gateway reserves the worst-case cost against the area’s budget; then debits the real cost and frees the reservation. Fully transactional, with row locks — with no budget, the call never leaves (HTTP 402).
Per-call cost ceiling, per-area budget, end-of-month forecast and automatic switching to a cheaper model of equivalent quality. Semantic cache (embeddings-based) eliminates repeated calls.
spend this month · Engineering area
$4,280 / 6,000
PII redacted before persisting · async
Every spend is traceable by area, account and token. Async logs (90d retention) with PII redacted (national IDs, email, phone, card) before persisting. Z-score detection suspends anomalous tokens — without blocking the response.
A provider error never becomes a wrong charge or a corrupted response: the gateway rolls back the reservation and returns 502 predictably. Rate limit (sliding window) and IP allowlist protect every token, always fail-closed.
library · 18 prompts
+ New promptExecutive summary
General · admin
Email {{tone}}
Marketing · you
Meeting minutes
General · admin
Support reply {{case}}
CX · john
{{…}} variables pre-fill the chat · scope by approval
Register reusable prompts with {{…}} variables and use them in the chat pre-filled with one click. Anyone can create; share as private (just you) or public for the whole company — with admin approval. Search by area and author.
The conversation doesn’t stop at text. With one click, the user turns what they discussed into a presentation (PPTX), document (PDF) or infographic — the model generates the structure and a server-side renderer builds the file, ready to download.
4 infographic styles · real product samples
// drop-in, OpenAI-compatible
The API is compatible with the OpenAI format. Point your SDK at the gateway, use a gw_… token instead of the provider key, and every call becomes authorized, metered and audited — without touching your code.
one line changes · governance on every call
// security and compliance
Provider keys encrypted with AES-256. Master key in an environment variable, never in the database. Revealable only by the admin, once, under confirmation.
The audit trail is append-only, chained by SHA-256. Deletion is blocked at the database. The chain is verifiable at any time.
Personal data (national IDs, email, phone, card) is redacted before persisting and before sending to the provider — closing off inbound and outbound leaks.
Any doubt about quota, authentication or state results in a refusal. Rate limit, IP allowlist and anomaly suspension protect every token.
Login via Entra ID / Microsoft 365, Okta or Google Workspace, with account provisioning on first access and role-based profiles.
Docker Compose stack as the production artifact: on-premise or in your cloud. Datastores on an isolated network, daily backup with configurable retention.
// frequently asked questions
No. The API is OpenAI-compatible and routes to OpenAI, Anthropic, Google or self-hosted models (vLLM, Ollama). Switching providers is configuration, not a code rewrite.
Overhead is minimal: observability runs in the background (it never blocks the response) and the semantic cache answers instantly on a hit. Reserve and settle are millisecond database operations.
Yes. The production artifact is a Docker Compose stack — on-premise or in your cloud, with datastores on an isolated network, daily backup and assisted zero-downtime migration.
PII (national IDs, email, phone, card) is redacted before persisting and before sending to the provider. With self-hosted models, data never leaves your network. Every operation lands in a verifiable audit trail.
Change your SDK base URL and use a gw_ token instead of the provider key. Works with Python, Node, cURL and automations like n8n — without changing your application.
Yes. Each automation (n8n flow, agent, script) is a headless identity with its own owner, budget, models and token — governed and auditable individually.
Book a demo and watch reservation, settlement and audit happening on a real call — in your environment.