§ Multi-model gateway · OpenRouter-backed · SOC 2 in progress
One workspace for 50+ AI models. The router picks the right model for each prompt and instantly retries another if one goes down — so requests never fail. Send a prompt to several models at once, track spend per user and project, and get one bill at month's end.
For teams shipping AI features who are done juggling provider keys, dashboards, and invoices.
§ 01 / PROVIDERS
Live status across every model graph we pull from — refreshed at the edge.
§ 02 / ROUTER
Decide which model at request time.
Stop hard-coding model names into your app. The router scores the prompt, picks the cheapest model that can answer, and cascades to a fallback if the primary stalls.
Cookie-signed JWT → BFF auth() → backendStream() proxies with X-Internal-Token and X-User-Id.
Regex + heuristic gate: research · coding · image · doc-analysis · reasoning · simple. Returns a single intent + a confidence band.
Plan-rank gate, entitlements check, daily/monthly token budgets. The cheapest acceptable model that matches the intent wins.
One retry on the same model, then current_model.fallback_model_id. 401/403 skip the retry — bad keys fail loudly.
start · delta · metadata · error · fallback · done. Every attempt logged to usage_logs; every failure to error_logs.
§ 03 / SWARM
Six specialists. One supervisor.
Fan a single brief out to a topology of specialist agents. The supervisor watches each return, resolves conflicts, and writes a single synthesized answer back to the stream.
§ 04 / COMPARE
Same prompt. Different minds.
Fan a prompt out to two or more models in parallel. We render both streams as they arrive, count tokens, and surface the cost delta so you can pick on evidence — not on vibes.
§ 05 / SURFACE
The tools the model can touch.
Three primitives that turn a chat box into a workspace — ingest a file, ingest a URL, rewrite a prompt. Each one is exposed as a first-class tool the router can call.
PDF · DOCX · CSV · PPT · MD. Chunked, embedded with pgvector, cited inline.
Pull HTML, strip chrome, follow citations, render a structured note with sources.
Rewrites vague prompts using established prompt-engineering primitives — role, constraints, format.
§ 06 / SPEC
A datasheet, not a brochure.
The full feature surface, named in the same terms the backend uses. If the system does it, it's listed. If it doesn't, it isn't.
Every prompt is scored for what it actually needs — research, coding, analysis. The cheapest model that can handle it wins, so you're not paying frontier prices for simple work.
If a model errors, we retry it once, then automatically switch to a backup. Requests don't fail just because one provider is having a bad day.
Responses stream in word by word as the model generates them — no waiting for the full answer to load.
Every user's data is isolated at the database layer. No user can ever see another's prompts, files, or history.
Set daily and monthly spend limits per user or project — enforced before each request. Every call and error is logged for a full audit trail.
Premium models gate by plan rank. Feature flags govern compare, council, search, file-chat, output-studio. No surprises at checkout.
Bundle chats, uploaded docs, and a system prompt into a Project. Every new conversation inherits the surface area you've already built.
Take any response into a structured editor — tone passes, length passes, format passes — then export markdown, html, or docx.
§ 07 / PLANS
one time · no expiry
billed annually
billed annually
scoped per deployment
All plans include token accounting, plan-rank model gating, and per-request logging. Free tier is lifetime — no card, no renewal. Paid plans: annual billed once, cancel anytime, no pro-rated clawback.
§ 08 / FAQ
Anything else, write us at info@azelaai.com — we read every line.
Yes. AzelaAI lets you run GPT-4o, Claude 4.5 Sonnet, Gemini 2.5, DeepSeek, Mistral, Llama and 50+ other frontier models inside a single workspace. Send one prompt to all of them at once with Compare Mode, switch between them mid-conversation, or let the smart-router pick the best model for the task. One login, one bill, every model.
§ 09 / END OF SHEET
Open an account, point a prompt at the router, watch every frontier model take a swing. Free tier needs no card — Pro is ready when the work catches up.
free tier · 10 messages total · no credit card