Two lines in your settings send Claude Code through the Kottos AI gateway to Anthropic. Your login or API key passes through unchanged. We never see a prompt and never store a credential. What you get back is the record: per request, which model, which venue, how long to the first token, how many tokens were cache reads, and the net dollars per million tokens you paid, model by model. That number is not on any price page.
Add the env block to ~/.claude/settings.json (or to the project's .claude/settings.local.json to scope it to one repository). The key below is the shared trial key: it is public on purpose and it only labels your traffic as the trial. Your Anthropic credential is the one Claude Code already holds.
// ~/.claude/settings.json { "env": { "ANTHROPIC_BASE_URL": "https://api.kottos.ai", "ANTHROPIC_CUSTOM_HEADERS": "x-kottos-key: kb_live_sPWBYzmGGOIqkCPUPbAfxsSwj6K-8iWJ" } }
Same file on every platform: ~/.claude/settings.json on Linux and macOS, %USERPROFILE%\.claude\settings.json on Windows, and the env block is identical in all three. Claude Code reads these two variables itself, so no shell profile is involved.
Prefer the shell? The same two as environment variables: export ANTHROPIC_BASE_URL=https://api.kottos.ai and export ANTHROPIC_CUSTOM_HEADERS="x-kottos-key: kb_live_sPWBYzmGGOIqkCPUPbAfxsSwj6K-8iWJ". On PowerShell: $env:ANTHROPIC_BASE_URL="https://api.kottos.ai" and the same for the header. Remove the two lines and Claude Code talks to Anthropic directly again; nothing else changed.
Then start Claude Code and open data.kottos.ai/trial. Your requests appear as they complete.
Anthropic's price page lists Claude Fable 5.1 at $10 per million input tokens and $50 output, with cache reads at $0.25. A Claude Code session is mostly cache reads, so what a session pays per million tokens sits far below the list price, and it moves with how well the cache is holding. The dashboard leads with that number for every Anthropic model, computed from the token counts Anthropic itself returns on each response: input, output, cache reads, and cache writes split by lifetime.
In plus out, after cache reads and writes are priced at their own rates. The number to compare models on, because it is the one you are billed.
Measured at the gateway, per request, with the provider's own queue and generation separated from the microseconds that are ours.
Every request that did not come back whole, with its status. Rate limits and 5xx are the fills that never arrived.
The trial dashboard is shared and public: everyone on the trial key lands on the same page, pooled. Rows carry a time, a model, a venue, a size and a cost, and nothing that names a person or a company; the one client-set field that could (the run label) is dropped from this view. A private account with its own dashboard and the full window is what the beta adds.
The record written for each request is a fixed 384-byte layout: timing stamps, model, venue, token counts, status. It has no field for prompt text and no field for the credential, so neither can be stored by accident. The open-source half of that layout is public in the llmbridge repository; read the struct instead of a privacy policy. Your Anthropic credential crosses the gateway in memory, is scrubbed from the buffer it arrived in, and reaches Anthropic over TLS.
A measurement. The gateway forwards your request to Anthropic byte for byte and records what came back. It does not change the model you asked for, does not pick another venue, and does not add tokens.
An account. The key is shared and may be rotated, and this page always carries the current one. There is no quota beyond what your own Anthropic plan allows, because the tokens are billed to you, not to us. For a private account, write to us.
Yes. Claude Code sends its own login in the Authorization header, the gateway passes it through unchanged, and Anthropic answers as it would have. The founder runs his own sessions this way.
The gateway adds microseconds. The hop to it adds your round trip to our host, which depends on where you sit. The first token is set by Anthropic's queue and generation either way. The dashboard shows both, separately.
Because a trial that needs a signup is not a trial, and the key spends nothing: your requests are billed by Anthropic to your own plan. The key labels traffic. It does not authorize money.
Delete the two lines. Claude Code goes back to talking to Anthropic directly, and nothing on your machine remembers the trial.
Questions and answers live in a public thread on GitHub, so the next person finds them. Posting needs a GitHub login. Nothing on this page is private, so keep keys and prompts out of it.
No widget? Ad blockers sometimes stop it. The same thread is at github.com/kottos-ai/community/discussions.