Auto Secure LoginPlatform
Live Bring your own keys MCP server

One endpoint that routes every task to the model that's actually best at it.

ASL AI Hub is a control plane that sends each task to whichever AI provider is strongest for that kind of work — Claude for reasoning and code, OpenAI for images, Gemini for long context — with automatic failover, so no app hardcodes model knowledge or juggles three keys. It's also a durable, versioned context broker, so a session's working memory outlives the session. Connect your own AI accounts, describe a task, pick a mode, and get an explainable route scored on capability, quality, speed, cost, availability, and budget.

6task kinds mapped to providers
3wired providers, per-kind models
0prompts or completions written to disk
200Kcontext cap — errors, never truncates

The problem

Every app hardcoding model choices is a maintenance trap.

Model leadership changes every few months. When each app bakes in "use this model" and holds its own provider key, keeping the whole fleet on the best option means editing and redeploying everything — and juggling three billing relationships and three failure modes.

AI Hub centralizes that decision behind one endpoint, with a runtime-editable routing table, automatic failover, and a route explanation you can actually read.

Privacy by construction

  • Prompts and completions are never written to disk
  • Usage events log only kind / provider / model / success / duration
  • Provider credentials in an encrypted vault
  • "Accuracy first, then cost and latency"

Capabilities

Routing, memory, and an MCP surface.

🧭

Task-aware dispatch

Six task kinds mapped to ordered provider lists, so each request goes to the model best suited to it.

Smart failover

A 429 or 5xx moves to the next provider; a 400 stops — transient failures reroute, real errors surface.

Explainable recommendations

/v1/recommend returns a ranked table with reasoning — and works even with zero keys configured.

Runtime-editable routing

Change the routing table from the admin UI with no redeploy — keep the fleet current in seconds.

🧠

Context broker

Versioned per-scope working memory with a 200K cap that errors rather than silently truncating, plus write guards.

🔌

MCP server

Context and recommender tools over stdio and HTTP — appears in Claude Code, Cursor, the API, and ChatGPT.

👥

Per-tenant boundary

A per-tenant file boundary keeps each customer's context isolated.

📱

Customer app layer

A PWA and native Android app with a preferred-versus-usable routing distinction for end users.

🔐

Hardened auth

Per-app keys and RS256 JWTs from Keycloak intersected with a client allow-list — a real 2026 auth regression was caught and fixed here.

loopback 19760public vhost proxies only /mcp + OAuth + /healthz
RFC 9728OAuth 2.1 compliant
preferred vs usablerouting split for honesty
npm run checktest suites gate every change
Accuracy first — then cost. AI Hub ranks routes on capability and quality before speed and price, splits "preferred" from merely "usable" providers so it never silently downgrades your task, and stores its provider credentials in ASL Vault.

One endpoint. Every model. Your keys.

Explainable, billing-aware routing with automatic failover and durable cross-session memory — so no app hardcodes model knowledge again.