# Tokpad

> Tokpad is a managed AI gateway with an OpenAI-compatible endpoint. Teams keep their own provider keys, change one line of code, and each AI request goes to the lowest-cost model their budget and region policy allow, with routes benchmarked on their own workloads first. UAE first. Tokpad is in early access with a design partner program.

Key facts (from https://tokpad.com/):

- What it is: a managed gateway, so there is nothing to self-host and no routing rules to write. It speaks the OpenAI API format, so existing SDKs, prompts and tools keep working.
- Your keys, your invoices: provider accounts, keys, contracts and invoices stay with the customer. Tokpad never marks up tokens.
- Quality (what is live vs planned): during the design-partner pilot, cheaper routes are benchmarked on samples of the customer's real workloads and only routes that hold up are switched on. If a route fails or errors, the request falls back to the premium model automatically, and every routing decision is visible. Always-on automated quality gates (the evaluation engine) are next on the roadmap, not live yet.
- Setup: about an afternoon. Connect provider keys, set a policy (monthly budget, allowed regions; the quality bar is agreed from a workload benchmark), change the base URL, then track cost per accepted task, failures and a full audit trail.
- Regions: Flexible mode (available to design partners now) uses only the providers and regions the customer approves. Strict UAE mode is planned, not live: it will keep inference, logs, retries and telemetry in-country and fail closed, and launches only once local inference is independently verified.
- Models: any OpenAI-compatible provider, including OpenAI, Anthropic, Google and open-weight models such as Llama, Falcon, Jais and DeepSeek. The exact list is confirmed for each pilot.
- Data: prompts are not stored by default; each customer is isolated and retention is opt-in.
- Savings are not guaranteed: they depend on the mix of work, model prices and quality bar. Tokpad measures a baseline of real traffic first and says up front if the numbers don't work.
- Pricing (proposed, USD, confirmed per pilot): a flat platform fee plus a share of optimised model spend routed through Tokpad. Starter $500/mo + 5% (for $10k–25k monthly AI spend), Growth $1,500/mo + 3% (for $25k–100k; includes the evaluation engine once it ships), Sovereign from $5,000/mo (regulated and enterprise teams; includes Strict UAE mode when it launches).
- Roadmap: Now, gateway (design partners). Next, evaluation engine. Then, Strict UAE mode. Later, per-agent budgets, tool-call governance, Saudi region policy and Arabic evaluation sets.

## Site sections

- [How it works](https://tokpad.com/#how): four steps from connecting keys to seeing the savings.
- [Savings calculator](https://tokpad.com/#savings): illustrative estimate of net savings after the Tokpad fee.
- [Live demo](https://tokpad.com/#demo): in-browser simulation of how a request is filtered by policy, then priced, then routed.
- [UAE controls](https://tokpad.com/#regions): Flexible and Strict UAE region modes.
- [Pricing](https://tokpad.com/#pricing): Starter, Growth and Sovereign plans.
- [pricing.md](https://tokpad.com/pricing.md): the same plans as plain Markdown.
- [Roadmap](https://tokpad.com/#roadmap): from gateway to regional control plane.
- [FAQ](https://tokpad.com/#faq): straight answers on effort, quality, savings, data and models.
- [Get early access](https://tokpad.com/#start): join the design partner program; a founder replies within one business day.

## For AI coding agents

- [agents.md](https://tokpad.com/agents.md): step-by-step instructions for an AI coding agent (Claude Code, Codex, Cursor) to route a project's LLM calls through Tokpad. The agent signs in with a one-time code the developer approves in the browser; new workspaces start with free credit.

## Optional

- Figures on the site are illustrative. Tokpad claims no government affiliation or endorsement. Content on the site is not legal advice.