Meerada — the live AI model exchange · LLManager (meerada CLI, package handover)

LLManager imports your Claude Code / Claude.ai / ChatGPT history and lets you continue it on any model — switch mid-conversation and the whole history moves with you. Arena grades models hourly on verifiable tasks and prices them by cost per verified task. Site: https://meerada.app/ · Download: https://github.com/ravemm-hub/meerada/releases/latest

🔎 Cross-check — the model that examines models: send one task to several models; every answer is examined by a jury of three models from the other labs (never its own family) — score, agreement, flagged claims, a winner. In LLManager (one button) and the terminal (meerada crosscheck "…" --models a,b,c). Verdicts stay on your machine and become the training set for Meerada's own judge. Spec: docs/JUDGE.md

🛡 Meerada Guard (in the same download): a watchdog on every model call — tells stuck from working hard, alerts in seconds when a task stalls or burns, and cages what leaves the workspace (keys, dotfiles, unknown hosts). Works beside Claude Code with zero setup (meerada guard), inside LLManager on every call, and company-wide from one policy file. Spec: docs/GUARD.md · page: https://meerada.app/guard.html

Measure what your LLM spend actually buys.

Live site: https://meerada.app/ · Hosted LLManager: deploy it in ~10 min →

One click reads render.yaml. Then set GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRET and OAUTH_REDIRECT_URI (the session/vault secrets are generated for you). Full walkthrough — including a local dry-run with no Google needed — in docs/DEPLOY.md.

Every dashboard shows you tokens and latency. Nobody shows you the only number that matters: CPAT — cost per accepted task — what one verified, completed unit of work really costs, including the retries, the reasoning-token burn, and the attempts that died on the way.

meerada (working name — may change) reads your LLM traffic locally, verifies task success with programmatic signals (test exit codes, schema validation, output contracts), and renders a ranked model leaderboard for your workload — not a synthetic benchmark.

Quickstart

pip install -e .
meerada record events.jsonl # ingest a trace source (metadata extracted locally)
meerada report              # ranked HTML leaderboard: CPAT, waste, success rates

Zero infrastructure: local SQLite, one self-contained HTML file, no accounts.

Privacy, by architecture

• Request/response content never leaves your environment — traces store only metadata, salted SHA-256 fingerprints, and aggregate scores.
• The salt is generated locally per database and is never transmitted.
• The report is a local file. Nothing is published anywhere.

What the numbers mean

Metric	Definition
CPAT	Σ cost of all attempts ÷ verified successful tasks
TTAT	Σ wall time of all attempts ÷ successful tasks
Attempts/win	Σ attempts ÷ successful tasks
Waste	retries + excess reasoning + uncached context + dead tasks

Every proportion carries its sample size and a Wilson 95% confidence interval. Every estimated (non-measured) number is labelled derived. Tasks with no verification signal are excluded, never guessed.

Status

P0 (core pipeline + report) complete. P1 in progress: task clustering, output-contract extraction, golden sets, and meerada pack — the handover pack that lets you migrate a workload between models with a measured gap report.

See SPEC.md for the full design. Development: make test, make lint.