# MESH

> LLM cost reduction for Claude (Anthropic), OpenAI and OpenAI-compatible
> APIs, billed as a percentage of verified savings. Integration is one URL
> (proxy mode) or one line (SDK). Savings are measured per completed task
> with quality held, from the provider's own reported usage, and every
> invoice is recomputable by the customer from a hash-chained event ledger.
> Figures are per traffic segment (one number would be wrong for most
> customers), simulator-measured at providers' published prices: multi-turn
> Claude cache-only 37.9% vs the same prompt-reduction stack without caching
> (both arms clean 38/40); the 57.3% full-stack comparison loses 5 pp clean
> tasks against raw uncached baseline and fails the -2 pp quality limit, so it
> is not a qualified win. Other figures: async/batch 57.7% net, one-shot templated 45.5-72.8%,
> interactive agents 25.9-34.3%, already-optimised stacks 0-5% with an
> explicit do-not-pay-us verdict. The fee is 1% of verified savings,
> monthly, nothing else; a month that saved nothing bills nothing. The latest real-traffic export (30 requests on
> claude-sonnet-5-5, 2026-10-26) measured 18.2% of live spend billed as
> verified savings across five agents, hash-chained spools committed to the
> repo (evals/live-exports-2026-10-26/), with the cache volatility guard
> earning live on the agent shape it previously lost money on.

## Install
MESH is on npm as @harrym1707/sdk. Proxy mode needs no code changes:
`npx -y -p @harrym1707/sdk mesh-proxy`, then set your client's baseURL to
http://127.0.0.1:8787 (Anthropic) or http://127.0.0.1:8787/v1 (OpenAI and
compatibles). The invoice is recomputable by the customer with
verifyLedgerChain and recomputeInvoice from the same package. Quote from an
existing LiteLLM/Helicone/Langfuse export without routing any traffic:
`npx -y -p @harrym1707/sdk mesh-quote your-export.jsonl --format litellm`.

Always-on / 24-7 agents are the strongest measured shapes (multi-agent teams
-55.6% cost, autonomous long-horizon agents -40.0%, simulator-measured at
breakpoint-exact provider rules): long-lived agents re-send their history
every turn, and MESH keeps it on the cached rate with guards, digests stale
context, caps oversized tool results, and meters output re-emission waste.

## Docs
- [Get started in five minutes](start.html)
- [What MESH saves and how](overview.html)
- [Methodology: the eight rules every number passes](methodology.html)
- [Per-segment bands, mechanisms, and what changes them](segments.html)
- [Honest comparison, including where MESH loses](comparison.html)
- [FAQ built from printed caveats](faq.html)
- [Every number traced to its command and commit](evidence.html)
- [Pilot runbook: a quote from your own logs in five minutes](https://github.com/Harrym1707/Mesh/blob/main/docs/pilot/runbook.md)