The number on your invoice is not the number
Most AI SaaS charges you per token. The number on your invoice is not the number the model provider charged. It is the provider's number plus a markup, plus a margin, plus a platform fee, plus a rounding convenience. You never see the real cost. You never see the real model. You never see the real request. The vendor sits between you and the weights, takes a cut on every token, and decides which models you are allowed to call. We think that shape is bad for users and bad for the long-term trust of the industry.
The markup problem
When a vendor resells model access, they add a margin. The margin is rarely disclosed. Industry numbers we have seen range from 1.5x to 3x the underlying provider's price. You pay the markup on every token, every request, every retry. You also lose model choice. The vendor decides which models you can call. They decide which model is default. They decide when to upgrade, when to deprecate, and when to silently swap the model behind a name you trusted. Your prompts tune themselves to a model you did not pick.
The markup problem has a second-order effect. Because the vendor is in the middle, they have to meter every request. They build rate limits. They build quota games. They build credits that expire, that round down, that do not transfer between models. The platform becomes an accounting layer between you and the model. Every feature in that layer costs you money and costs the vendor engineering time. None of it improves the model. All of it insulates you from the real cost curve, which is the only signal that would let you choose wisely.
How BYO key works in Tangison Agent
Tangison Agent uses OpenRouter as the default model router. The flow is simple. You create an OpenRouter account. You fund it with the budget you choose. You paste the API key into your .env file. The agent reads the key, calls OpenRouter, OpenRouter routes to the model you specified, the model returns a response, and the agent writes it to your workspace. Tangison never sees the key, the bill, or the response.
Tangison never sees the key. The key lives in your .env, on your server, and is never sent to us. Tangison never sees the bill. OpenRouter bills you directly, with the provider's real price on every line. Tangison never marks up the usage. We do not sit between you and the model. We do not meter your requests. We do not rate-limit you. We do not have a credits system. If you want to switch routers tomorrow, you change one environment variable and the agent keeps running.
The agent configuration specifies which model to call per task. You can run Claude for nuanced reasoning, GPT for code, Mistral for cost-sensitive bulk, Llama for offline work, and Qwen for multilingual, all in the same agent, all on your own key. Swap models per task, per environment, per user. The agent does not care. The skills, the channels, and the audit log all stay identical across the swap.
What you gain
First, model choice. You are not locked into one vendor's roadmap. When Anthropic ships a new Claude, you call it. When OpenAI ships a new GPT, you call it. When Mistral, Llama, or Qwen leapfrog, you call them. The agent's skill surface stays the same. Only the model changes, and only when you decide. You can A/B test models on real workloads against your real evals, not the vendor's marketing evals.
Second, price transparency. Your OpenRouter dashboard shows exactly what each request cost. You can attribute spend per task, per user, per project. You can set budgets and alerts at the OpenRouter level. You can wire that data into your own observability stack. The number on your invoice is the number the model provider charged, full stop. No markup, no platform fee, no rounding convenience.
Third, no surprise bills. There is no Tangison metering layer that can drift. There is no Tangison rate limit that can throttle you mid-launch. There is no Tangison quota that can expire. Your spend is bounded by your OpenRouter budget, full stop. If a runaway loop sends 10,000 requests at 03:00, the OpenRouter budget caps it before your invoice does.
Fourth, no quota games. You are not buying credits that round down. You are not losing value when a model is deprecated. You are not migrating credit balances between platforms. You pay for what you use, when you use it, at the provider's price. Fifth, the ability to swap models per task. Reasoning tasks can use a stronger model. Classification tasks can use a cheaper one. Bulk transcription can use a fast one. The agent config file is the only thing that changes.
What we gain
A sustainable business model that does not depend on reselling your compute. Tangison Agent is free and MIT licensed. The Supported plan, for teams that want Tangi on call for setup, upgrades, and incident response, is a flat monthly fee. The Enterprise plan, for organizations that want custom skills, private channels, and an SLA, is an annual contract. Neither plan resells model access. Neither plan marks up your tokens. Neither plan meters your requests.
This matters because it aligns our incentives. We do not make more money when you use more tokens. We make more money when you stay on the Supported plan year over year. The way we keep you is by shipping a better agent, better skills, better docs, and better channels. Not by trapping your spend, not by rounding your credits down, not by silent model swaps that make you consume more. The business model is the moat, not the metering layer.
Closing
Getting started takes about an hour. Clone the repo. Run the install script. Drop your OpenRouter key into .env. Open the dashboard. Send your first message. The docs walk through every step with copy-paste commands and working examples. If something breaks, file an issue on GitHub and we will look at it.
If you want a hand on the first deploy, message Tangi on WhatsApp. Tell him what you want to automate, which models you want to call, and where you want to host. He replies within one business day, Monday to Friday, 09:00 to 17:00 CAT. Tangison Agent is built by Tangison, an independent software brand founded by Tangi Iigonda. The parent brand publishes notes, talks, and product work at tangison.com.