← meerada.appthe Meerada group

How it works, in three short walkthroughs

No sign-up needed to follow along. Each walkthrough plays on its own; click a step to jump.

📊 Reading the exchange meerada.app — every model, measured hourly

1Every row is a model. CPAT = cost per verified task: what you actually pay for one correct answer, not the list price per token.
2Grades come from real tasks with checkable answers, re-run every hour. A grade is never a vibe, always a pass/fail count.
3Grey numbers are derived (estimated from list price); green and red are measured. Only measured rows can rank.
4Filter by task type or budget, then compare two models side by side — the gap report tells you what you'd save and where quality breaks.
modelgradeCPATlist $/M
gemini-2.5-flashA · 46/48$0.004$0.30
llama-3.3-70bB · 39/48$0.006$0.59
qwen3-32bB · 37/48$0.005$0.40
mistral-smallderived~$0.008$0.20
💡 CPAT: the price of one correct answer. Cheap tokens that fail often are expensive.
taskexpectedgemini-2.5-flash
JSON extract #143 fields✓ pass
Hebrew date parse2026-09-15✓ pass
SQL fix #7row count 12✗ fail
Unit test gen4/4 green✓ pass
🕐 Re-run every hour · 48 tasks · pass/fail only, no opinions
modelgradeCPAT
gemini-2.5-flashA · measured$0.004
mistral-smallderived~$0.008
gpt-5-miniderived~$0.012
⬜ grey = no key for that lab yet → estimated from list price, cannot rank
gemini-2.5-flashvsllama-3.3-70b
metricflashllama
grade46/4839/48
CPAT$0.004$0.006
breaks on—SQL, long JSON
📄 Handshake report: move 80% of this workload → save 33%, keep SQL on flash

🎛️ LLManager one app for every model, on your own keys

1Import your history: Claude Code, Claude.ai or ChatGPT export → drag the file in. Every conversation lands in one library.
2Open any conversation and keep talking — the whole history is sent along, so the new model knows everything the old one knew.
3Switch models mid-conversation (Handshake). Pick by the exchange's grade and CPAT, not by brand.
4Run many models on many tasks in parallel and keep the answer that passed. Your keys, your machine, no subscription.
⬇ drop claude-export.json / conversations.json
📚 Library: 312 conversations · 2 sources
✅ imported in 4 s — nothing leaves your machine
Continue the migration plan we drafted yesterday
Picking up from step 3 (the RLS policies)… here's the next migration file:
🧠 the full thread travels with you — no re-explaining
Model: claude-sonnet gemini-2.5-flash ★ A · $0.004 gpt-5-mini
🤝 Handshake: switched to gemini-2.5-flash. History: 41 messages carried over. Continuing…
💱 picked by grade and CPAT from the exchange
taskflashllamaqwen
write tests✓✓✗
fix bug #212✓✗✓
summarize PR✓✓✓
⚡ 3 models × 3 tasks in parallel · keep what passed

🛡 Guard the watchdog for AI at work

1Guard watches your AI sessions (Claude Code, agents, batch jobs) and learns what "working" looks like: tokens per minute, files touched, progress.
2A task that stalls — same output looping, no files changing, tokens still burning — is flagged in seconds, not at the end of the bill.
3You get one alert with the numbers: what it cost so far, what it would cost if left alone, and a stop button.
4Stuck vs. working, told apart. The $10-in-one-task session that started Guard never happens again.
session build-tribes-b132 · 14 min
tokens/min 2.1k · files changed 7 · tests green
🟢 working — normal rhythm learned over 3 sessions
session refactor-auth · 31 min
tokens/min 9.8k · files changed 0 for 9 min · output repeating
🟠 looping — same edit attempted 6×
session refactor-auth
🛑 Stalled task · spent $2.40 · projected $11.20 if left alone · Stop · Nudge
📣 one alert, with numbers, before the bill
todaysessionsstalls caughtsaved
Tue62$18.70
🧾 stuck vs. working, told apart — every day