Your company runs AI models in Microsoft’s cloud, and developers use them through tools on their laptops. The problem: out of the box there’s no way to give each developer their own key, their own monthly budget, or their own usage meter — it’s one shared pool, and one person can drain it for everyone. FoundryGate fixes that. Each developer automatically gets a personal key and a monthly AI budget; every request is counted; when the budget runs out, requests stop instantly; and asking for more is one click to request, one click to approve. Budgets come in a few set sizes — a standard one, a bigger one for heavy users, and unlimited — so “more” means moving up a size, not haggling over a number.
The rest of this page explains the same idea in more technical terms. For the next level of detail, see the system overview, or the whole system in one picture.
Here is the entire life of a single request, from the developer’s terminal to the model and back. The tier numbers are read out of the infrastructure at build time, so they are whatever the deployment actually enforces.
…and one of three answers comes back, decided right there:
Azure AI Foundry is excellent at what it does: hosting model deployments and exposing them through Azure API Management. But when you move from a single team’s prototype to an org-wide AI platform, you hit a wall — Foundry was designed for deployment management, not developer access governance.
When you set up Azure AI Foundry with APIM today, you get:
That works fine when you have three engineers sharing a dev environment. It breaks when you have 200 developers across 12 teams, each expecting a fair share of a limited TPM budget.
Per-developer token budgets
Azure AI Foundry has no concept of “alex@company.com gets 5 million tokens in June.” FoundryGate assigns each developer a monthly budget tier — Standard, Power or Unlimited — enforced at the APIM subscription level, not as a soft warning.
Individual APIM subscription keys
Instead of a shared key or RBAC access, every developer gets their own APIM subscription key. Usage is attributed per person. Keys can be rotated or revoked per person without affecting anyone else.
Entra-synced access
Access follows your directory. Being given access to FoundryGate in Entra is what creates a developer’s account and issues their key; losing it is what revokes them. Team groups are synced separately — moving somebody into a team’s Entra group moves their budget to that team’s, without anyone editing a spreadsheet.
Self-service for developers
Developers sign in with their Entra account, see their current quota and usage, rotate their own key, and request a quota increase with a justification — all without contacting IT.
| Capability | Azure AI Foundry alone | With FoundryGate |
|---|---|---|
| Per-developer token budget | ❌ Deployment-level only | ✅ Per user, per month |
| Individual API keys | ❌ Shared key or RBAC | ✅ One APIM subscription per developer |
| Hard quota enforcement | ❌ No blocking mechanism | ✅ Real-time 403 at the gateway when the monthly budget is spent |
| Per-developer burst control | ❌ One developer can saturate the deployment | ✅ Per-developer TPM cap → targeted 429 + Retry-After |
| Entra-based auto-provisioning | ❌ Manual RBAC assignment | ✅ Assigned in Entra → account + key |
| Team budgets from the directory | ❌ Nothing to sync | ✅ Entra group membership → that team’s budget tier |
| Automatic deprovisioning | ❌ Manual RBAC removal | ✅ Access removed in Entra → key deleted |
| Monthly reset | ❌ Manual | ✅ Azure Function runs at 00:01 UTC on the day you set (the 1st by default) |
| Quota increase workflow | ❌ None | ✅ Submit → admin approves → quota applied |
| Per-developer usage visibility | ❌ Aggregate only | ✅ Dashboard + audit log per user |
| Model deployment management | ❌ Portal only | ✅ Claude models are set up by the infrastructure deploy; admins manage OpenAI model deployments from the FoundryGate UI |
| Which models a team may use | ❌ Everything on the deployment, or nothing | ✅ Admins choose which models each budget tier is allowed to use, from the FoundryGate UI — anything else is refused at the gateway |
| Deployment-level rate limits (TPM/RPM) | ⚠️ Shared ceiling, 429s passed to callers | ⚠️ Mitigated — backend pools + circuit breakers reroute around saturated deployments; the platform ceiling itself remains |
FoundryGate sets up the gateway (Azure API Management) and the AI model deployments (Azure AI Foundry) for you, inside your own Azure subscription, and then governs access to them.
FoundryGate tracks usage · enforces budgets · manages the key lifecycle
Key provisioning: When a developer first signs in, FoundryGate calls the APIM Management API to create an APIM subscription under the developer’s budget tier (Standard, Power or Unlimited). The key is issued to the developer. Foundry never knows about individual developers — APIM does.
Quota enforcement: The gateway itself enforces each developer’s budget in real time. A budget is one of the configured tiers — the tier product’s policy carries the monthly token quota, which is why quotas come in set sizes rather than arbitrary numbers. APIM’s llm-token-limit policy, keyed on the developer’s subscription, carries their monthly token quota (requests return 403 the moment it’s spent — no polling lag) and a per-minute token cap (bursts get 429 + Retry-After, which coding CLIs back off on automatically). Budgets reset with the UTC calendar month on their own. A reconciliation Function keeps the dashboard’s usage figures authoritative from the gateway’s request logs.
Cost: the dashboard can show an estimated cost per developer, from a rate card an admin fills in at /config — Azure bills Claude as one aggregate meter, so nobody but you can split it per person. It is a reporting figure, always labelled an estimate, and absent entirely until those prices are configured. The gateway enforces tokens, never dollars.
Entra sync: FoundryGate reads your Entra directory through the Microsoft Graph API using the app’s own identity — no passwords or secrets to manage. There are two separate syncs, and it is worth keeping them apart:
Both syncs run on demand and are safe to repeat; running one twice with nothing changed in the directory does nothing the second time.
FoundryGate is for platform and infrastructure teams who want to:
