AIFab puts one gateway in front of every language model, and builds assistants, knowledge bases and workflows on top of it, with MCP tools wired in. It installs inside your network.
Teams reach every model through the gateway, build on it in the workspace and connect their own systems as tools.
OpenAI, Anthropic, Gemini, Bedrock, Mistral and your own GPU servers behind one address, with keys, quotas, masking, routing and metering.
Upload documents into knowledge bases, combine them with a model and tools into an assistant, and chain steps into workflows with human approval.
Register MCP servers once and let assistants use your databases, ticket systems and internal services, with every argument scanned and every credential sealed.
Applications keep their OpenAI or Anthropic SDK and change only the address and the key. The rest happens in the gateway, in this order.
Keys are matched as keyed HMAC-SHA256 digests. The plain key is never stored.
National IDs, tax numbers, IBANs and 16 other identifier kinds are masked before the call leaves.
Requests, tokens, concurrency and budget are checked together, in one atomic step.
To an outside provider or to the least-loaded of your own replicas, only among models the key is granted.
A usage event and a double-entry ledger posting, written to disk before the model is even called.

Every assistant call goes through the same keys, limits, masking and audit as any other request.

MCP servers are registered in the platform, not in each application.

Built for Turkish and Gulf data, and applied while the answer streams back, not after.

When capacity is short, quotas and weighted classes decide who waits, and batch jobs cannot starve interactive users.

The offline package carries its images, a software bill of materials for each one and a signed manifest.
Chat Completions, Responses, Messages, embeddings, rerank and OCR, translated to whatever the upstream model speaks.
OIDC, SAML 2.0, LDAP and SCIM 2.0 provisioning, with TOTP, passkeys and recovery codes as second factors.
Each audit record carries the SHA-256 of the one before it, so an exported chain can be checked on its own.
Usage is on disk before the model call, and the ledger only accepts new entries.
Every query names its tenant, row-level security backs it up, and caches are partitioned per user.
Sensitive configuration changes wait for a second person's approval.
If its configuration goes stale, its quota store goes away or its scanner stops answering, the gateway refuses new work.
Traces and logs carry ids, codes and timings, never prompt text, keys or client addresses.