With just two line changes in your Claude Code settings, you can send your requests through Kottos AI and evaluate them. You do not need to change your login or API key, and we never log or store prompt and responses text. With this trial you can observe the most recent records per request showing model, venue, how much time to first token, how many tokens were cache reads, and the net dollars per million token at the current rates for each model. The trial account is public so you may see the requests metadata from other users too.
For this trial, you only need to update ANTHROPIC_BASE_URL and ANTHROPIC_CUSTOM_HEADERS in Claude Code settings. The first variable needs to point your requests to the Kottos AI gateway, and the second needs to contain the token that gives you permission to use the trial account. Note that we have throttles in place as a safety measure.
Open ~/.claude/settings.json (or the project's .claude/settings.local.json) and add the env block below to it:
// ~/.claude/settings.json { "env": { "ANTHROPIC_BASE_URL": "https://api.kottos.ai", "ANTHROPIC_CUSTOM_HEADERS": "x-kottos-key: kb_live_sPWBYzmGGOIqkCPUPbAfxsSwj6K-8iWJ" } }
The settings filename is always the same but location may depend on the platform: for Linux or macOS ~/.claude/settings.json, and for Windows %USERPROFILE%\.claude\settings.json on Windows.
If you prefer a terminal, set the two as environment variables: export ANTHROPIC_BASE_URL=https://api.kottos.ai and export ANTHROPIC_CUSTOM_HEADERS="x-kottos-key: kb_live_sPWBYzmGGOIqkCPUPbAfxsSwj6K-8iWJ" for Linux or macOS, and $env:ANTHROPIC_BASE_URL="https://api.kottos.ai" and $env:ANTHROPIC_CUSTOM_HEADERS="x-kottos-key: kb_live_sPWBYzmGGOIqkCPUPbAfxsSwj6K-8iWJ" for Windows.
Claude Code picks the config changes automatically so there is no need to restart. After you add the env block, check the dashboard data.kottos.ai/trial and your requests should appear as they complete. If you want to stop the trial, simply remove the env block.
There are many aspects of optimal routing we would like to show but an image is worth more than a thousand words. That is why we set up the trial so you can draw some conclusions by looking at basic metrics. We encourage you to get an intuition of how the LLM inference infrastructure works by testing the trial and reach out to us if you would like to have a dedicated account.
The net effective rate is an important number to consider when deciding on a model or provider. This is what you are billed on.
The gateway measures the timing at different steps of the request path so we understand your timing bottlenecks. You can analyze one of them which is the time to first token.
If a request does not come back or is partially completed, you can also see its status. If you hit rate limits or 5xx errors, those are examples of prompts not completed by the provider.
The trial dashboard is shared and public so everyone on the trial will share resources. We do not show identifying data.
The record for each request is a fixed byte layout that does not allow for prompt or responses text, nor credentials. Credential headers are also scrubbed after each request is processed. Additionally, all connections run with TLS.
A measurement. We forward your request to Anthropic byte for byte and record what comes back. There is no change, i.e. the model does not change, the venue is always anthropic direct, and tokens are the same.
An account. The trial is public and the key is shared. We may rotate the key if and when needed but this page will always show the current one. Your quotas are Anthropic so everything is what your plan allows. Nothing is billed 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, and Anthropic answers normally. We run our own sessions this way.
The gateway adds microseconds to the request path. The only noticeable latency could be the trip latency added by going through the gateway server. This is capped at 120 ms within the US. This number is still small compared to TTFT and also depends on where you send the request from. If this is something that matters for your work, reach out to us because this is exactly the type of problems Kottos AI optimizies for.
We believe that a trial should take seconds to setup and not require a sign-up.
Delete the two lines. Once these are removed, Claude Code goes back to talking to Anthropic directly. There is no need to restart Claude Code either.
We created a public thread on Github for questions or comments you may have. This is public to allow anyone to read previous discussions, so do not post any keys or private prompts.
If a blocker stops the widget above, see github.com/kottos-ai/community/discussions.