You bring the models and the provider contracts. The gateway routes, meters and checks every request inside your perimeter.
Each team calls its own provider. Nobody can say who asked which model, and for which purpose.
Costs land in separate provider accounts. Finance sees the bill after the money is spent.
A prompt can go to any region the provider chooses. Nothing stops it at the boundary.
Calls an OpenAI-compatible endpoint. Change one base URL. Keep the code.
Identifies the caller, checks the budget and the policy, then records the request.
The model you chose, under your contract, in the region you approve.
Anar supplies no models. The gateway governs the models you already use.
Each application gets a virtual key that a team owns. Teams never see the provider key.
Budgets and rate limits per customer, team and key. Cost is counted on each request.
Routing rules on content and caller. Content scan through Anar Guard.
Each request is logged with payload, cost and latency. Admin actions go to an audit trail.
The gateway returns an error to the caller. It also blocks failover to a provider in another region.
Shipped. Turn it on per policy.Today the check covers chat requests that match a residency policy. Embeddings, audio and images are on the roadmap.
Response to the caller
data_residency_violation
Fallback to another region: blocked.
- One static binary with the console built in. Nothing else to install.
- No call to Anar at run time. No licence server.
- Your database: a database that you run.
- Provider keys from your vault or from the environment.
| Item | Status |
|---|---|
| Runs without root rights | Shipped |
| Large payloads to your object store | Shipped |
| Signed offline bundle with SBOM | Roadmap |
native provider integrations, plus any endpoint that speaks the OpenAI API.
Native provider integrations, November 2026.
Arabic-aware routing: a rule can send Arabic prompts to the Arabic-capable model you licensed.
added at the median, with no optional features turned on
added at the median, with the full governance chain
requests per second on one process
added to a 200 ms first token
Measured September 2026 on a shared development machine: 8 cores, loopback, a mock provider, median of 3 runs. We can run the same test on your hardware.
Customers contain teams. Teams own keys. A key inherits each limit above it.
ShippedAnar Gateway console, virtual keys. Demo data.
See volume, success rate, latency, tokens and spend. Open any request to see who sent it and what it cost.
ShippedAnar Gateway console, request log. Demo data.
| Capability | Status | Note |
|---|---|---|
| Routing, budgets, keys, request log | Shipped | On in every install. |
| Residency refusal, cost export | Shipped | Turn on per policy. |
| Role-based access | Partial | Four roles set in the configuration file. Off by default. |
| Audit trail | Partial | Records admin actions. No export or tamper evidence yet. |
| Content scan | Partial | Needs Anar Guard. Passes traffic if Guard is down. |
| SSO, SCIM, alerts, PII redaction | Roadmap | Designed. Not built. |
- Routing, budgets and keys
- Request log and metrics
- Residency refusal
- Runs inside your network, with no call home
- Sign-in with your identity provider
- Roles you manage in the console
- Tamper-evident audit trail
- Signed offline bundle with SBOM
- Arabic console
- PII redaction and alerts
Scans prompts and answers for unsafe content and personal data.
Scores your AI controls against GCC and EU frameworks.
They can send their model calls through the gateway when you connect them.
Identity, budgets, routing policy and the request record, under every product above.
A one-month pilot: we deploy in your environment, connect your providers and run it with one team.
anarlabs.ai/contact