Connect your agent · Kottos AI

Set up

Every app starts the same way: add the server, sign in with your email, and ask Set me up with Kottos AI. The agent sees how far your account is and does only what is left, so a second app finds the account and the token the first one made.

Claude Code

In a terminal, add the server once, for every project, and sign in:

claude mcp add --transport http --scope user kottos-account https://account.kottos.ai/mcp
claude mcp login kottos-account

Start Claude Code with claude, ask Set me up with Kottos AI, and approve each step. It creates your account with $10 of credit for Kottos AI fees, asks you to accept the terms of service and the privacy policy, writes a token to ~/.kottos/token on your computer, readable only by you, and registers only its SHA-256 hash.
Within about a minute the token is live, and Claude Code adds Kottos AI to its settings. It needs no restart: your next message is your first request through Kottos AI. Ask Show my recent requests to see it arrive.

Claude Code in VS Code uses the same server and settings; sign in there with /mcp in its chat panel. Your Anthropic login or key stays where it is: Claude Code sends it with each request, and Kottos AI passes it through and never stores it.

Codex

In a terminal, add the server and sign in. codex mcp add usually opens the sign-in itself; codex mcp login opens it when it did not, or signs in again.
```
codex mcp add kottos-account --url https://account.kottos.ai/mcp
codex mcp login kottos-account
```
Start Codex with codex and ask Set me up with Kottos AI. It creates your account or finds it, makes the token or finds the one another app made, and writes a Kottos AI profile, ~/.codex/kottos.config.toml, leaving your own configuration as it was.
With your OpenAI API key in OPENAI_API_KEY, codex -p kottos sends Codex's requests through Kottos AI. Signed in with ChatGPT instead, Codex keeps its requests on OpenAI and still has your account's tools.

The Codex IDE extension and the ChatGPT desktop app read the same Codex configuration, so they have the server and its sign-in too. The Kottos AI profile is for the codex command line.

Claude or ChatGPT chat

Add https://account.kottos.ai/mcp as a custom connector with OAuth sign-in. In Claude: Customize › Connectors › Add custom connector, and let it register automatically. In ChatGPT: chatgpt.com/plugins, then + and Add custom MCP server, with OAuth, created as a plugin; then type @ in a new conversation and pick it. ChatGPT's help says full MCP is for Business and Enterprise plans and Pro gets reading tools only; there, create your account on your account page and ask the chat about the rest.
Ask Set me up with Kottos AI. A chat runs on its maker's servers, so it creates your account but cannot make a token on your computer: create it with one click on your account page, unless Claude Code or Codex made it already.
Then ask about your spend, savings and prices from any chat, on the web, desktop or phone.

Then ask it

Click a question to copy it, then paste it into your agent or chat.

Tools

whoami, create_account and set_token: your account, its balance and token, an account with its starting credit, and a token made on your computer.
get_started and list_routes: the setup for Claude Code, Codex or a chat, the settings and code from your account page for a first request to Anthropic, OpenAI or Amazon Bedrock, or for any route the gateway offers for a model.
recent_requests and usage: your last ten requests, and your requests, costs and Kottos AI fees for today, 7 or 30 days.
simulate: what your own past requests would have cost on another model and venue, from your account page's simulator.
list_models and compare_venues: list prices for each venue and service tier that sells a model, and time to first token from a probe that runs every minute.

Your token and keys

Your Kottos AI token is made on your computer and never passes through the chat or reaches us: we keep only its SHA-256 hash, as for a token from your account page. Your agent saves it in ~/.kottos/token, the file .kottos/token in your home folder, readable only by you, and tells you so; we email you too. Apps on that computer read it from there. On another computer, copy that file to the same place and keep it private, or create a new token on your account page, which replaces it. Replacing a token and adding credit stay on your account page, which can also create your first token if you would rather copy it yourself. Your provider keys pass through the gateway with each request and are never stored.

What the numbers are

Prices are list prices collected from providers' public pages. They can be out of date or wrong, and your contract can differ, so check the provider's page before you buy.
The probe sends one short, fixed prompt to each measured offer once a minute from Ashburn, Virginia, through the Kottos AI gateway. Its time to first token is mostly the network trip and the provider's queue, which is less than what a long prompt waits. Not every offer is probed, and the tool says which are.

Limits

120 requests a minute from one address. Claude's and ChatGPT's own servers are exempt, since every one of their users calls from them.
We have tested Claude Code. Codex and the chats are new, so tell us at [email protected] if a step differs.
Protocol versions 2026-07-28, 2025-11-25, 2025-06-18 and 2025-03-26. Each request gets one JSON reply, and there is no session.

Privacy

The server answers from your own account's records, and what it returns goes to your agent and, through it, to the AI provider your agent runs on. We count requests per day by client name, method, tool and the model matched, and store no IP address and nothing else a request sends. The privacy policy has the details.

Prices without an account

The price and probe tools also run with no sign-in at https://data.kottos.ai/mcp, the server listed in Claude's connectors directory.

Questions

Kottos AI, Inc. · [email protected]