One assistant on top of every model — frontier and open-weight. Routel picks the right one for each task, double-checks its answers before you see them, remembers what matters, and keeps your data yours. Typically ~80% cheaper than running everything on a frontier model.
We lived all of these. Routel exists to fix them.
Four wedges, one assistant.
Every request is classified and sent to the cheapest model that can handle it — a fraction-of-a-cent open model for routine work, a frontier model only for genuinely hard reasoning. You see the cost and the routing reason on every reply.
For multi-step goals, Routel runs plan → act → verify → repair. An independent checker model must approve the result before you see it. If it fails, Routel auto-repairs and re-checks — so you stop paying for broken attempts.
Preferences and facts live in a portable file you can read, edit, and export. Your assistant remembers your stack, your tone, your constraints — and you can take it with you anytime.
Bring your own key: prompts run on your provider account. We redact emails, phone numbers, keys, and card/SSN patterns before any cloud call, and we never train on your data.
Routel is built by a small, independent team of engineers and operators who got tired of watching AI bills balloon and agents fall over on real work. We’re not a frontier lab and we’re not trying to be — we don’t train giant models. We build the layer that makes the models you already pay for cheaper, more reliable, and more private.
We believe the best AI product isn’t the one with the biggest model — it’s the one that uses the right model for each job, proves its work, and treats your data and your money with respect.
In 2025–2026 the tools we relied on changed billing overnight and sent developers surprise invoices. Agent products burned credits on retries that failed silently. Frontier subscriptions charged premium prices for tasks a model 30× cheaper could nail. And our prompts — often containing sensitive code and customer data — disappeared into systems we couldn’t inspect.
So we built the assistant we wanted: one that routes intelligently, checks itself, caps your spend with your permission, and keeps your keys and data under your control. Routel is a client you point at your own provider accounts — like an editor or a terminal — not a reseller standing between you and the model. That design is deliberate: it keeps you in a direct, compliant relationship with each AI provider.
Email and password. Accept the Acceptable Use Policy. No card required.
Paste an OpenRouter key (one key → many cheap + frontier models) or OpenAI/Anthropic. It’s encrypted at rest and used only for your requests.
Pick a monthly hard cap. Chat or run the agent — every reply shows the model, the cost, and what frontier would have cost. No surprises, ever.
Privacy isn’t a setting here — it’s the architecture.
Your provider key is encrypted with AES-256-GCM and used only to serve your requests. We never pool one key across users.
Emails, phone numbers, API keys, card numbers and SSNs are stripped before any off-device call — on by default.
We don’t use your prompts or outputs to train models, and we never build competing models from provider outputs.
Every request logs which model and provider served it, so you always know where your data went.
A hard monthly cap you control, with an alert at 80%. A surprise bill is structurally impossible.
Export your memory anytime as a plain JSON file. Your context is yours to keep or take elsewhere.
A small flat fee for the layer. Compute runs at cost on your own key — we don’t mark up tokens.
Prices shown are illustrative launch pricing. You’ll always see compute cost at cost; the flat fee is for the routing, verification, memory, and privacy layer.
No. Routel is a Bring-Your-Own-Key client: your requests run on your provider account under your terms with each AI provider — like an IDE or terminal that you point at your own credentials. We never pool one API key to resell access, which is both better for you and the compliant way to build.
Through an OpenRouter key you reach dozens of models — DeepSeek, Llama, Gemini, GPT, Claude and more — with one key. You can also connect OpenAI or Anthropic directly. Routel routes each task to the cheapest one that can handle it.
Routine work (summaries, formatting, simple Q&A) goes to models that cost a fraction of a cent; frontier models are reserved for genuinely hard reasoning. On a typical mix that’s around an 80% cut versus running everything on a frontier model — and you see it live, per message.
The agent loop adds an independent verification step that catches incomplete or off-task results and repairs them before delivery. It reduces errors significantly — but no AI is perfect, so always review important outputs. AI output carries no warranty.
Yes. Routel is an installable web app (PWA): open it on your phone and choose “Add to Home Screen,” or click “Install” on desktop. No app store needed.
No. We don’t use your prompts or outputs for training, and we don’t build competing models from provider outputs. Your memory file is yours to export anytime.