FAQ — built from the caveats we already print

How do I install MESH? One command: npx -y -p @harrym1707/sdk mesh-proxy starts a local proxy with a hash-chained ledger — then set your client's baseURL to http://127.0.0.1:8787 (Anthropic) or http://127.0.0.1:8787/v1 (OpenAI). No signup, no account, and your API key passes through without being stored. The full walk-through, including recomputing the invoice yourself, is on Get started.

Does MESH work with my language or framework? Yes, if it can set a base URL — which is Python, TypeScript, Go, curl, LangChain, LlamaIndex, and effectively everything else. Proxy mode is language-agnostic by construction; the one-line SDK mode (withMesh(client)) is TypeScript/JavaScript today. It covers Claude (Anthropic), OpenAI, and any OpenAI-compatible endpoint.

My agents run 24/7 — what does MESH actually do for always-on workloads? Long-lived agents re-send their whole history every turn, so they re-buy their own memory forever — on real agent traffic, each output token gets re-read ~27 more times downstream. MESH keeps that history on the provider's cached rate (with guards that refuse placements that would lose money), digests stale context byte-stably, caps oversized tool results behind a quality gate, flags byte-identical retries, and meters output re-emission on your statement. The measured always-on shapes are our strongest bands: see savings by segment.

Can MESH break my app? The default levers are lossless: compacted schemas decode identically, removed duplicates are exact, prefix repair reorders text without deleting it, and the reducer never throws — any failure returns your request untouched. Lossy levers (output caps, description trimming, tool-result capping) are off until a replay gate proves on your own recorded requests that quality held.

Why do you quote me a range and not one number? Because savings depend on your traffic's shape. The quote comes from your own export and names your segment; a vendor quoting everyone one number is wrong for most of them. Sometimes our verdict is literally doNotPayUs.

Is 57.7% (async) real? It is a faithful measurement of a realistic simulation at providers' published prices, net of assumed discount-tier fallback rates, with quality held at 38/40 clean tasks — and it is labelled simulator-measured everywhere it appears. Your pilot replaces it with your own number before any fee exists.

You're billing a counterfactual — how is that dispute-proof? The same way energy performance contracts have done it since 1995: a defined baseline, provider-verbatim measurements, and an adjustment method agreed in advance. Concretely: every request writes an append-only hash-chained event; the invoice is a pure function over those events; you can re-run it and must get our number. Savings you already had are itemised and excluded. Negative savings (e.g. a cache-write premium that didn't pay off) reduce the bill — they are never clamped to zero.

Do I have to change my code? One line or none. SDK mode wraps your existing client (withMesh(client)). Proxy mode needs no code at all: run the MESH proxy locally and point your client's base_url at it — your API key passes through verbatim (MESH never holds credentials), responses stream back untouched, and every request appends the same hash-chained ledger event the SDK writes. A stream whose usage cannot be read is recorded *unpriced*, never guessed.

What do you do with my prompts? Analysis runs locally; prompts never leave your machine. The ledger records usage numbers, lever names, hashes and verdict references — not prompt text. One optional exception you control: the SDK can POST usage metadata (never prompt text) to an endpoint you configure yourself; it is off unless you set it. Full detail: data handling.

What if the provider changes prices? The price catalogue is dated and source-linked. A new rate is a new dated entry; history is never silently repriced, and an uncovered date is "unpriced", which excludes the row rather than guessing.

What does it cost? 1% of verified savings, monthly — nothing else. No subscription, no seats, no minimum. A month that saved nothing bills nothing, and negative lever months count against us.

What's the catch? Two, published: busy interactive agents cap around 29% because the provider's cache already does the heavy lifting; the 34.3% sparse figure awaits a per-model cache-minimum correction and may come down for some models. (A third is now closed: Anthropic batch pricing was uncatalogued; as of 2026-10-27 it is catalogued at the provider's definitive 50% discount, so Claude async traffic prices natively instead of being OpenAI-priced.) When a number moves for honesty reasons, this site moves with it — the async headline already went from 64.3% to 57.7% that way.