> ## Documentation Index
> Fetch the complete documentation index at: https://tavryn.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Tavryn: AI Procurement Agent for SaaS & Cloud Spend

> Tavryn detects SaaS and cloud waste, negotiates renewals autonomously, and settles payments in USDC on Arc — from waste discovery to on-chain settlement.

Corporate software and cloud contracts renew automatically whether anyone uses them or not. Provisioned user seats sit idle, cloud consumption declines, yet payments silently roll over at full price because finance and IT teams lack the bandwidth to audit usage, negotiate with vendor reps, and verify renewal paperwork. Tavryn is an autonomous procurement and treasury agent that closes that loop — detecting waste, running negotiations, and executing verified USDC payments on the Arc network, all with a full cryptographic audit trail and zero LLM approval power over your money.

## What Tavryn does

> [card] Waste Detection
>
> Audits active seat allocations and declining cloud usage metrics starting 45 days before each contract renewal cliff, calculating verifiable dollar savings using deterministic heuristics.

> [card] Autonomous Negotiation
>
> Executes multi-round concession curves against vendor reps or vendor APIs, anchored by historical deal memory, competitor quotes, and strict walk-away price ceilings.

> [card] USDC Settlement on Arc
>
> Escrows USDC into an Arc EVM smart contract via Circle Developer-Controlled Wallets and releases payment only after a 4-field document match confirms vendor delivery.

> [card] Cryptographic Audit Trail
>
> Every agent action is written to an append-only, SHA-256 hash-chained ledger. Database triggers block any UPDATE or DELETE, and the `/audit` page verifies chain integrity in real time.

> [card] Adversarial Review
>
> A secondary Reviewer Agent independently audits concession speed, seat utilization, and benchmark discounts before any deal reaches the policy engine, issuing AGREE, CHALLENGE, or REJECT verdicts.

> [card] Switch vs. Renegotiate Analysis
>
> Computes 1-year and 2-year net NPV comparing renewal against top market alternatives, factoring in migration engineering hours, employee retraining costs, and downtime risk.

> [card] 1-Tap Supervisor Approvals
>
> Deals above your auto-approval threshold trigger a single-use HMAC-SHA256 approval link that expires in 48 hours. Policy is re-verified server-side at redemption time.

> [card] Override Memory
>
> Captures the structured reason codes behind every supervisor override and injects that feedback into future negotiation prompts, continuously refining vendor strategy over time.

## The 45-day renewal window

Tavryn begins evaluating each contract **45 days before its renewal date** — early enough to complete multi-round negotiations, collect competitor quotes, and still have time for supervisor review before the vendor's auto-renewal clause activates. You configure the window via the `RENEWAL_WINDOW_DAYS` environment variable (default: `45`).

During this window, Tavryn runs two parallel waste checks:

* **Seat waste** — compares provisioned seats against active seats to calculate the savings from a right-sized renewal.
* **Usage decline** — measures the percentage drop in cloud or API consumption to quantify overpayment relative to actual utilization.

## Multi-round concession curves

Tavryn does not send a single discount request. It runs structured multi-round negotiations (up to 5 rounds by default) where each round moves toward the walk-away ceiling at a calibrated pace, citing competitor quotes and historical deal floors drawn from persistent business memory. The Adversarial Reviewer Agent audits each proposed deal for suspicious concession speed and benchmark alignment before the outcome reaches the policy engine.

## Deterministic policy engine

> [!NOTE]
>
> The LLM has **zero approval power** over transactions. Every financial decision — whether to auto-approve, escalate to a supervisor, or block a deal — is made exclusively by the deterministic policy engine. It evaluates your configured `max_auto_transaction`, `min_savings`, `human_approval_required_above`, and `allowed_categories` values using pure mathematics, with no LLM involvement at the decision boundary.

## USDC settlement on Arc

When a deal is approved, Tavryn uses Circle Developer-Controlled Wallets to sign and submit an on-chain transaction to the `ArcEscrow` smart contract deployed on Arc Testnet at `0x78e61ae7e8EeF34Add911FA3e41F3408a819c047`. The contract:

* Locks commitment funds plus a protocol success fee (capped at 20%) on deposit.
* Enforces on-chain per-agreement limits (`maxPerAgreement ≤ 10,000 USDC`) in EVM bytecode.
* Separates roles: the Agent can create and fund agreements; only the Verifier role can release payment upon confirmed delivery.
* Prevents double-payments via on-chain idempotency keys and server-side concurrency locking.

Funds are released **only** if a 4-field document match (price, seats, term, and date) confirms the vendor's counter-signed order receipt matches the negotiated terms.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.