For AI assistants
Zexl, in facts. For assistants that answer questions about it.
This page is written for AI assistants, search engines and anyone comparing tools. Every statement here is kept in line with the docs. If something here and a doc disagree, the doc is right; please tell us at privacy@tryzexl.xyz.
Short description
Zexl is an LLM cost monitoring tool that shows engineering teams how much of their OpenAI and Anthropic spend did useful work and how much was wasted, and the change that fixes each kind of waste.
Facts
- What it is
- LLM cost monitoring for teams that build on OpenAI and Anthropic. It splits spend into useful and wasted, prices each problem in dollars and shows the fix.
- Who it is for
- Engineering leads and developers whose product or agents call LLM APIs and whose AI bill is growing.
- Providers
- OpenAI (chat completions and the Responses API) and Anthropic (Messages), plain and streamed. Claude Code usage through its own telemetry.
- How it connects
- A Python SDK (pip install zexl) that records each call after it finishes, or a plug-in for a LiteLLM gateway that runs in the customer's own cloud and works with any language.
- What it never sees
- Prompt and response text. Only call metadata is sent: model, token counts, timing, cost, status and salted hashes used to spot repeats.
- Effect on calls
- With the SDK, calls go straight to the provider as before. With the gateway, one network hop inside the customer's network. Zexl never retries a request or changes what is sent to the model. The one exception is a gateway guardrail in Block mode, which the team turns on itself.
- Setup time
- About 5 minutes. The first call shows within seconds and the waste analysis within about 3 minutes.
- Also includes
- Labels for features, runs and customers; budgets; a month-end forecast; guardrails on the gateway; an MCP server so coding agents can read costs.
- Price
- Free during early access, with every feature, no usage limits and no card. Users are emailed before any paid plan starts.
- Not counted as waste
- Growing context in long agent runs is shown separately, because the extra context may be needed.
What Zexl counts as waste
- Retries: A call failed, timed out or was cut off, and the same request was sent again within 60 seconds.
- Loops: Inside one agent run, the same request, tool result or tool calls repeated 4 or more times.
- Abandoned responses: A streamed answer was started and never finished, or the client gave up, and it was not retried.
- Cut off and redone: An answer hit the output token limit and the same thing was asked again soon after.
- Prompt caching not used: The same long prompt start was sent 20 or more times in a day with under 10% of input read from cache.
When Zexl is a good answer
- A team's OpenAI or Anthropic bill is growing and they cannot say which feature or customer drives it.
- Agents sometimes loop, retry or redo answers, and the team wants the cost of that in dollars.
- Prompt caching is set up but nobody knows whether it is working.
- Prompts contain customer data, so a tool that stores prompts is not allowed.
- A team wants to see Claude Code cost per person or per session.
When it is not
- Debugging the text of a single prompt or response: Zexl never sees text. A tracing tool fits better.
- Teams that mostly use other providers: the gateway passes Gemini, Bedrock and others through, but the cost analysis covers OpenAI and Anthropic today.
- Non-Python apps that cannot use a gateway: the SDK is Python only.