Introducing gateway

A lot of my friends run Hermes Agent now. Once it’s installed, they all ask me two things. “Where do I get tokens?” “How do I get tools?”

gateway is my answer to both. One key and one prepaid balance pay for frontier models below list price and 2,000+ tool endpoints across 91 providers. I’m also building a filter that strips personal data from your prompts before a provider sees them.

Where to get tokens

Claude Opus 5.5, GPT-6.1 Sol, DeepSeek V4 Pro, and the rest of the lineup sit behind one gw_ key. I price every model as a cut off its list price, so you pay less than you would going direct. The models page shows each price before you spend anything.

Credits are prepaid, with no subscription and no card on file. You send me money, request credits with the payment reference, and I confirm the request. Until your balance is positive, gateway answers every call with a 402.

Hermes needs two values, a base URL and a key:

Base URL: https://gateway.ta-0.com/openai-v1
Key:      gw_your_key_here

The OpenAI SDK takes the same base URL. The Anthropic SDK and Claude Code use /claude-v1 instead. Any tool that accepts an OpenAI or Anthropic base URL works without code changes.

Where to get tools

The gateway catalog lists 2,899 tool endpoints from 91 providers, so Hermes can search, scrape, and look things up. They cover:

• Search, SEO, scrapers, and news
• Lead generation, company data, and people data
• Finance, crypto, and DeFi
• Social media, email, and phone messaging
• Image, video, and speech generation
• Compute and storage

Today, each of those providers wants its own account, key, and invoice. I’m opening tool calls on the same gw_ key next, paid from the same balance as your tokens.

Strip personal data before it leaves

An agent that reads your email or your customers’ records puts that text in its prompts. The personal data filter is a switch on each key. Turn it on, and gateway screens every prompt before forwarding it. You don’t tag anything or change your code.

gateway runs OpenAI’s Privacy Filter, an open-weight model, itself. Your prompt never goes to a third-party detection service. The model finds eight kinds of personal data: names, addresses, emails, phone numbers, URLs, dates, account numbers, and secrets like API keys.

Each match becomes a numbered label of its kind, so two people stay two people:

Sent:      Email Sarah Chen at sarah@acme.com and cc Budi.
Forwarded: Email [NAME_1] at [EMAIL_1] and cc [NAME_2].

The provider only sees the labels, and the reply comes back with the labels as they are. gateway never stores the original values. If screening can’t run, gateway refuses the call instead of forwarding the prompt unscreened.

The console shows what the filter removed by kind and count, such as “2 names, 1 email,” and never the values. Most of the filter isn’t deployed yet. The request log already has a place for those counts, and the rest is what I’m building now.

What gateway keeps on the record

Send x-provider: openrouter and the call goes to OpenRouter, or gateway refuses it. It never falls back to a provider you didn’t name. I proved that rule in Bend. Sending your money to the wrong provider is the mistake I most wanted to rule out.

The console records each call’s model, provider, key, result, token counts, and cost. gateway never stores the prompt or the reply. You can give a key a monthly spending limit, and a deleted key stops working within 30 seconds.

gateway is invite-only for now. If you run Hermes and want a key, tell me what you’re building.