Audric

Your money, handled.

Conversational finance on Sui. Save, pay, send, borrow — all by talking to your AI financial agent. Built on t2000 infrastructure.

Live at audric.ai

---

The five products

Audric is exactly five products. Everything you can do is one of them. (S.18 reverted S.17's Finance retirement — Intelligence was carrying both "the moat" and "the home for every financial verb," and Send/Receive overlapped Pay. Finance now owns save/credit/swap/charts; Pay owns send/receive.)

Product	Description
🪪 Audric Passport	Trust layer — sign in with Google, non-custodial Sui wallet in 3 seconds, every write taps to confirm, sponsored gas. Wraps every other product.
🧠 Audric Intelligence	The brain (the moat) — 4 systems: Agent Harness (26 tools), Reasoning Engine (12 guards), Memory (MemWal vector store), AdviceLog. Multi-step orchestration ships as 14 skills (markdown playbooks in @t2000/mcp baked into the published bundle, exposed to Cursor / Claude Desktop as MCP prompts). Engineering-facing brand; users experience it as "Audric just understood me."
💰 Audric Finance	Manage your money on Sui — Save (NAVI lend, 3–8% APY on USDC or USDsui — strategic exception added in v0.51.0), Credit (NAVI borrow USDC or USDsui against your savings, health factor visible — repay must use the same asset as the borrow), Swap (Cetus aggregator across 20+ DEXs, 0.1% fee), Charts (interactive yield/health/portfolio viz from chat). Every write taps to confirm via Passport.
💸 Audric Pay	Money primitive — send USDC to anyone, receive via payment links / QR. Free, global, instant on Sui. No bank, no borders, no fees.
🛒 Audric Store	Creator marketplace at audric.ai/username. Sell AI-generated music, art, ebooks in USDC. Coming soon (Phase 5).

How it works

1. Sign in with Google — no seed phrase, no keys, no downloads (zkLogin)
2. Fund your wallet — deposit USDC to your Sui address
3. Talk — tell Audric what you need

Your money lives in a non-custodial wallet. Audric executes transactions, but you approve every one. Zero gas fees.

Stack

Layer	Technology
Framework	Next.js 15 (App Router)
Auth	zkLogin via Enoki (Google OAuth → Sui wallet)
Gas	Enoki sponsored transactions (zero gas for users)
AI	@t2000/engine — 26 tools (18 read + 8 write), reasoning engine, extended thinking, canvas. Skills (multi-step playbooks) ship from @t2000/mcp.
Database	NeonDB (Prisma) — 15 models (users, user preferences, profiles, memories, financial context, advice log, conversation log, session usage, payments, watch addresses, linked wallets, portfolio snapshots, turn metrics, app events, service purchases)
Sessions	Upstash Redis (KV)
Styling	Tailwind CSS v4, Agentic Design System
Hosting	Vercel

Audric Intelligence — the 4-system moat

Not a chatbot. A financial agent. Four systems work together to understand your money, reason about decisions, and get smarter over time. Every action still waits on Audric Passport's tap-to-confirm.

🎛️ Agent Harness — 26 tools, one agent

18 read tools + 8 write tools covering balance checks, savings (save_deposit, withdraw), lending (borrow, repay_debt), swaps (swap_quote, swap_execute), payments (send_transfer), payment links, on-chain analytics, BlockVision-backed pricing (token_prices), reward compounding (pending_rewards + harvest_rewards), and SuiNS resolution (resolve_suins). Read tools execute in parallel; write tools require user Passport confirmation and execute sequentially under a transaction mutex (TxMutex).

Post-S.245 + S.269 + S.277 (May 2026): Down from 37 tools. pay_api + mpp_services (S.245), save_contact + the 3 invoice tools (S.269), and the Volo trio + web_search + protocol_deep_dive (S.277 "Earns Its Keep" audit) were cut from the engine surface — paid-API commerce returns as Audric Store, invoicing collapses into payment links (label/memo encode invoice context), and Volo liquid staking remains available in @t2000/sdk + @t2000/cli + @t2000/mcp for non-Audric consumers.

⚡ Reasoning Engine — thinks before it acts

• Extended thinking — always-on for Sonnet/Opus (adaptive mode). Haiku for low-effort queries
• Adaptive effort — classifies each turn as low/medium/high/max and adjusts model + thinking depth
• Guard runner — 12 guards (10 pre-execution + 2 post-execution hints) across 3 priority tiers (Safety > Financial > UX) enforce balance freshness, health factor limits, slippage thresholds, irreversibility warnings, address-source / address-scope / asset-intent gates, swap preview confirmation, and retry protection
• Skills (multi-step playbooks) — 14 markdown playbooks shipped from @t2000/mcp, baked into the published npm bundle at build time and exposed to MCP clients (Cursor, Claude Desktop, claude-code CLI) as skill-<name> prompts. The 6 multi-step playbooks (t2000-rebalance, t2000-account-report, t2000-borrow with safe-borrow logic, t2000-withdraw with emergency-close logic, t2000-save with swap-and-save, t2000-send with contact-resolution) absorbed the orchestration that pre-Phase 6 lived in a YAML recipe runtime in @t2000/engine (deleted v0.7a Phase 6, May 2026). Skill content guides the LLM through multi-step intents; the engine just runs the tools the LLM picks.
• Preflight validation — input validation on send, swap, pay, borrow, and save before execution
• Prompt caching — static system prompt + tool definitions cached across turns for lower latency and cost

🧠 Memory — knows your finances (MemWal-backed)

Long-term memory lives in @mysten-incubation/memwal (Mysten vector memory). Every turn, the prepareStep hook calls memwal.recall(latestUserMessage) to fetch top-K relevant facts → injected as a <memory_recall> system-prompt block. After the turn, the onFinish callback calls memwal.analyze() to extract new facts (preferences, risk tolerance, goals, on-chain patterns) from the conversation. Plus a short-term daily on-chain orientation snapshot — UserFinancialContext (savings/wallet/debt USD, health factor, current APY, recent activity) — refreshed at 02:00 UTC and injected as a <financial_context> block. Both used silently to calibrate tone + recommendations — never surfaced as nudges.

v0.7d Phase 6 Block A (2026-05-21): the previous UserFinancialProfile (Claude-inferred preferences) + UserMemory (Jaccard-deduped facts) + ChainFact (7 classifier rows) Prisma tables were dropped. MemWal absorbs all three surfaces into a single vector store with organic fact extraction.

📓 AdviceLog — remembers what it told you

record_advice tool writes AdviceLog rows; buildAdviceContext() hydrates last 30 days into every turn so the chat doesn't contradict itself across sessions. Episodic memory (UserMemory) and the full conversation log run alongside it for the future self-hosted model migration.

What was deleted in the April 2026 simplification: Copilot suggestions, scheduled actions / DCA, morning briefings, rate alerts, auto-compound, the features-budget allowance, the proactive-nudges pipeline, savings-goal milestone celebrations, follow-up queues, and the proposal pipeline behind BehavioralPattern. zkLogin can't sign without user presence — "autonomous" was reminders dressed up as agency. Update (S.103, May 2026): the broader savings-goal layer was fully retired in SPEC 17 — SavingsGoal Prisma table, 4 savings_goal_* engine tools, GoalsPanel UI, settings tab, and openGoals snapshot field all dropped. The "track my savings progress" job is now served by health_check + portfolio_overview + yield_summary. See the S.0–S.19 + S.103 entries in audric-build-tracker.md for the locked decisions on what we will not bring back.

What shipped recently — Spec 1 + Spec 2

Two harness upgrades on top of the 4-system base:

Spec	Versions	What it added
Spec 1 — Correctness	@t2000/engine v0.41.0 → v0.50.3	attemptId UUID v4 stamped on every pending_action (stable join key from action → on-chain receipt → TurnMetrics.pendingActionOutcome row). modifiableFields registry — fields the user can edit on a confirm card without losing the LLM's reasoning. EngineConfig.onAutoExecuted hook so auto-permission writes (currently none in Audric) participate in the same telemetry as confirm-gated ones.
Spec 2 — Intelligence	@t2000/engine v0.47.0 → v0.54.1	BlockVision swap — replaced 7 defillama_* tools with one token_prices tool; balance_check + portfolio_analysis rewired to BlockVision Indexer REST. Sticky-positive cache + retry/circuit breaker for graceful 429 handling. <financial_context> orientation block injected at every engine boot from the daily 02:00 UTC UserFinancialContext snapshot — every chat starts oriented, no warm-up tool calls. attemptId-keyed resume so two pending actions in the same turn never clobber each other's outcome. S.277 (May 2026): protocol_deep_dive retired — engine no longer talks to DefiLlama; rates_info covers the in-product safety lens.

Specs are local working documents (AUDRIC_HARNESS_CORRECTNESS_SPEC_v1.3.md, AUDRIC_HARNESS_INTELLIGENCE_SPEC_v1.4.1.md). Cross-repo contracts live in audric/.cursor/rules/audric-transaction-flow.mdc, audric/.cursor/rules/write-tool-pending-action.mdc, and t2000/.cursor/rules/agent-harness-spec.mdc.

Canvas Visualizations

Interactive HTML canvases rendered in-chat: portfolio timeline, activity heatmap, spending breakdown, yield projector, health simulator, watch list, and full portfolio dashboard. Generated via the render_canvas tool and delivered as canvas SSE events.

Rich Cards

Structured card types for tool results: balance, savings, health, staking, protocol deep dive, price, swap quote, transaction history, receipts, spending, yield, activity summary, payment links, and invoices.

Additional Features

• Health Factor in chat — surfaced prominently in health_check and balance_check cards. As of S.31 (2026-04-29) there are zero proactive surfaces; HF is shown when the user asks (and chat naturally surfaces it whenever they touch credit)
• Session pre-fetch — balance + savings data injected at turn 0 for faster first responses
• Streaming tool dispatch — read-only tools fire mid-stream before the LLM finishes
• Tool result budgeting — large results auto-truncated with re-call hints
• Microcompact — deduplicates identical tool calls in conversation history
• Daily-free billing — 5 sessions per rolling 24h for unverified users, 20 for verified. No paywall, no allowance setup
• Unified data layer — centralized portfolio + activity modules consumed by all API routes, canvases, and engine context

Architecture

audric.ai (this repo)
├── @t2000/engine    ← Agent engine (26 tools, reasoning, MCP, streaming, skills via @t2000/mcp)
├── @t2000/sdk       ← Core SDK (wallet, balance, transactions)
├── @suimpp/mpp      ← MPP payment client (Sui USDC)
└── @mysten/sui      ← Sui blockchain client

t2000.ai (separate repo)
├── CLI, SDK, MCP server, gateway, contracts
└── Infrastructure that powers Audric

Development

pnpm install
pnpm dev              # Start dev server (Turbo + Next.js)
pnpm build            # Production build
pnpm typecheck        # TypeScript check
pnpm lint             # ESLint
pnpm test             # Vitest

Environment

Copy .env.example to .env.local and fill in the required values. See CLAUDE.md for the full list.

Brand

• Audric is the consumer product — what users see
• t2000 is the infrastructure — what developers build with
• suimpp is the protocol — the open payment standard

License

MIT