Live preview
Æ

AAIOS-TCM Console

multi-model cost & time router · live ledger
Workspace skills, history & tools are sandboxed to the active domain
Requests
Spent
Saved
Saved %
Cache hits
What do you want to know?
Ask in plain words. It picks the cheapest model good enough, and the receipt is in the ledger.
Enter to send · Shift+Enter for a new line
Chat
One chat, three efforts — ⚡ Standard smart-routes one model, 🎭 Council debates it, 🐝 Super Heavy runs the full swarm. Reply to steer; it auto-compacts like Claude.
What the switches do
  • Smart mode — a cheap model reads your question first, sets its difficulty & tier, and asks you to clarify if it's vague. Better routing, tiny overhead.
  • Compound — finds your best prior answer to a similar question and builds on it, so your answers compound and improve every time you ask. Free & local. (On by default.)
  • Web search — grounds the answer in live web results (Claude). Best for prices, versions, current events. Runs on Anthropic and isn't cached. (On by default.)
Smart mode
Compound
Web search
⌘/Ctrl+Enter
Registry

Models

ModelTier$·MTokHealth
Savings by reason (window)
Debate

AI Corner — models challenge each other

A panel of 2–3 models you choose each answer the same question, then challenge each other over several rounds — probing weaknesses, conceding strong points, revising — and a moderator synthesizes the strongest joint verdict. Every panelist can web-search and is seeded from your Library (Compound), both on automatically.

AI Corner vs Graph Debate — which do I use?

🗣 AI Corner — a roundtable

Diverse models argue from different angles and a moderator merges them. Symmetric: every model is a peer. You control who's on the panel and who moderates.

Best for open-ended, strategic, subjective or creative questions where multiple perspectives beat one — brainstorming, trade-offs, "what would you do".

⚖️ Graph Debate — an assembly line

Asymmetric roles: one model drafts, a different-provider model audits it for errors, the heaviest model arbitrates, and a Verifier fact-checks it against live sources.

Best for factual/technical questions where correctness matters most — research, code, analysis, anything you'll rely on.
load a key to see models…
each turn is a real routed, ledgered call — cost shows per bubble & in stats
Multi-agent

Graph Debate — create · audit · arbitrate · verify

The Super Heavy pipeline: several Architects draft in parallel from different expert stances, a Merge folds them into one super-draft, three specialist Critics (security · logic · performance) attack it, the Sandbox actually runs any code it wrote, the Arbiter writes the final, and the Verifier fact-checks it with a confidence rating. The tracker below lights up each stage live. Easy questions skip all of it and get one cheap pass.

What runs automatically
  • Library grounding (Compound) + live web search are on by default — the debate builds on your prior answers and checks current facts.
  • Models auto-assign (primary executor drafts + arbitrates, critics on a different provider). Pin either role below to override.
  • 🧪 Sandbox (E2B): if the answer contains Python, it is executed in an isolated cloud VM — a crash sends it back for repair before you ever see it.
  • ⚡ $0 replay: asking the identical question again serves the stored verified answer from cache — no tokens spent.
  • 🛡 Guard: a local safety model screens every prompt before any money is spent.
  • The Verifier adds a 🔎 confidence chip + sources, and every run is saved as a debate — 👍 the good ones distill into reusable Skills.
easy questions get one cheap pass · tick Force full debate to always run the whole swarm
Folders

Categories

History

All questions

Load a key, then ask something to start your library.
🧠 AI analysis of your library — patterns & draft skills (on demand)
an AI reads your Library, summarises patterns, and drafts reusable skills
Click Refresh to analyse your Library.
Documents

Document & Legal Dashboard

Upload leases, insurance, court filings, contracts, invoices, IDs. An AI pulls out the parties, dates, amounts and obligations, and links related documents into a case by their shared identifier (e.g. a court-case number). Files stay on your server.

Drop files here or
PDF or text · up to 15 MB each
only new files are processed — re-running is cheap
Economics

How the saving is computed

Every request is priced twice — at the model that actually answered, and at the baseline (claude-opus-4-8). The gap is your saving. Four levers create it:

1
Routing — a cheap classifier scores difficulty and sends each prompt to the cheapest model good enough (Haiku $1/$5 vs Opus $5/$25). Your 80%.
2
Exact cache — the identical prompt again returns instantly for $0.
3
Semantic cache — a reworded past question (any language) also returns for $0.
4
Batch lane — non-urgent work (deadline ≥ 1h) uses provider batch APIs at −50%.
Hands-on

See each lever fire

A
Ask a question in Playground → note the saved line (routing).
B
Send the same prompt againexact_hit, cost $0.
C
Reword it and send → semantic_hit, cost $0.
D
Turn on Smart mode → a cheap model sizes the question, escalates only when it truly needs a bigger model, and asks to clarify if it's vague.