Weckr

What is Weckr?

Weckr measures what each customer costs you in LLM calls and compares it to what they pay. It sits beside the request rather than in front of it: you wrap your existing provider client, and Weckr records tokens, model, latency and cost after the response returns.

How it integrates

Two added lines in TypeScript (npm i @weckr/sdk) or Python (pip install weckr-sdk), passing userId, feature and plan alongside your normal options. Calls go straight to the provider, so nothing is added to the critical path, and if Weckr is unreachable logging fails silently rather than blocking the call. OpenAI, Anthropic, Gemini and Kimi are supported directly, with guides for LangChain, CrewAI, Vercel AI SDK, Next.js and FastAPI. An MCP server exposes the same data to agents.

What it does with the numbers

Cost is grouped per user, per feature and per plan, with worst-margin customers sorted to the top. Spending caps are set per plan in the dashboard and enforced by the SDK before the call: on a hit it either throws WeckrCapError or silently downgrades to a cheaper model from the same provider. Velocity alerts fire on runaway agent loops (50,000 tokens in five minutes) via Slack webhook and email, with an optional weekly digest.

Data handling

Prompt and response text is never transmitted. The SDK sends metadata only: token counts, model, latency, and the identifiers you pass in. The full SDK is MIT licensed and public, so the claim is checkable.

Pricing

Hobby is free for up to 50,000 requests a month, with 90-day log retention and no card required. Pro is $49/month flat for up to 1M requests, adding spending caps, loop detection, model recommendations and unlimited retention, with a 7-day trial. Verified 31 August 2026.

Who it is for

Teams selling a subscription product with LLM features underneath, where a flat plan price meets variable per-user model cost. Narrower than a general observability platform: it answers margin questions rather than debugging traces, and stores no prompts to replay.