DROP-IN FOR CLAUDE CODE, CURSOR & CI

Sit in the request path.
Keep the receipts.

Every LLM call passes through one gateway that enforces policy before the provider is called, meters every dollar to a team, and signs a tamper-evident receipt of what happened. The audit-grade evidence layer for AI — one product, not five stitched together.

EARLY ACCESS · NO CREDIT CARD REQUIRED
ONE REQUEST · SIX STOPS · ZERO BLIND SPOTS--:--:-- UTC · ● LIVE
INGRESSidentified & keyed
POLICYenforced in-path
ROUTEany provider, one key
LEDGERevery $ posts to a team
EVIDENCEsigned receipt, hash-chained
OUTCOMEStied to business results
14:32:04alice.chenclaude-sonnet-4.5$0.41PASS ✓
14:31:58ci-backendclaude-haiku-4.5$0.06CACHE-HIT
14:31:51sales-assistclaude-sonnet-4.5$0.28REDACTED ◆
PII REDACTED IN-PATH 847× THIS MONTHCACHE RECOVERED $2,140 MTDRECEIPTS SIGNED 1.24M QTDINJECTIONS BLOCKED 23×COST / TASK ▼ 22% QOQCHAIN STATUS VERIFIEDPII REDACTED IN-PATH 847× THIS MONTHCACHE RECOVERED $2,140 MTDRECEIPTS SIGNED 1.24M QTDINJECTIONS BLOCKED 23×COST / TASK ▼ 22% QOQCHAIN STATUS VERIFIEDPII REDACTED IN-PATH 847× THIS MONTHCACHE RECOVERED $2,140 MTDRECEIPTS SIGNED 1.24M QTDINJECTIONS BLOCKED 23×COST / TASK ▼ 22% QOQCHAIN STATUS VERIFIED
01 — THE THESIS

Nobody else sits in the path and keeps the evidence.

Observability tools watch traffic after the fact. Guardrail services score risk but can't stop a request. FinOps platforms see the invoice, not the request. Developer-analytics tools count activity, not outcomes.

Wardin is the audit-grade evidence layer for AI — it enforces, meters, and signs from inside the request path, so a policy violation never reaches a provider, every token has an owner before finance asks, and every call leaves a tamper-evident receipt.

Regulation is turning that receipt from nice-to-have into required: the EU AI Act's high-risk logging and record-keeping duties land in December 2027 and August 2028 — the exact artifact a request-path gateway produces as a byproduct. It's the convergence the category Gartner calls AI TRiSM keeps buying in pieces, and nobody else ships in one path.

5 → 1
tools replaced by one control plane
0 ms
between violation and enforcement — it happens in-path
100%
of requests emit signed, audit-ready receipts
02 — ONE PATH, ONE RECORD

Bought separately everywhere else. Native here.

ENFORCE

Policy-enforcing gateway

PII redaction, injection shields, model allowlists, and budget caps execute before the provider is called. A violation never leaves your network — there is nothing to clean up afterward.

REPLACES: guardrail service + gateway
METER

Finance-grade cost allocation

Every dollar posts to a team, a virtual key, and a person — as it's spent, not when the invoice lands. Budgets enforce themselves; month-end closes clean into NetSuite.

REPLACES: FinOps platform + spreadsheets
ATTRIBUTE

AI-effectiveness evidence

The same signed receipt proves whether governed spend produced accepted work — sessions tied to merged-PR outcomes from GitHub. Raw token counts are excluded by design, so you can't game it by burning more.

REPLACES: developer-analytics tool
03 — EVIDENCE BY CONSTRUCTION

Every request leaves a receipt. Every receipt is signed.

Each call emits a tamper-evident record — actor, model, policy checks, cost — chained by hash to the one before it. Editing any past record breaks the chain, so an auditor verifies the evidence rather than taking it on trust.

Every check maps to the controls the EU AI Act and NIST AI RMF ask for. Pull a signed evidence bundle for the recent chain in one click, or let a background job carry it into a Wardin-managed, Object Lock WORM store for long-term retention — then hand an auditor an offline verifier that checks the whole chain without a Wardin account. The artifact those record-keeping rules demand — not a compliance certificate, the evidence underneath one.

ED25519
signed at emit time
HASH-CHAINED
tampering is self-evident
S3 · PARQUET
WORM cold tier, Object Lock
WARDIN / RECEIPT
REQUEST IDreq_8f2c41ab
TIMESTAMP14:32:04 UTC
ACTORalice.chen · prod-engineering
MODELclaude-sonnet-4.5
PII / SECRETS SCANPASS ✓
INJECTION SHIELDPASS ✓
BUDGET CHECKPASS ✓
MODEL ALLOWLISTPASS ✓
TOKENS3.1k / 812
DEBIT$0.41
PREV HASHc41d…9e0b
THIS HASH7fa2…b3d8
SIGNED · ED25519 · TAMPER-EVIDENT
04 — THE STACK YOU WON'T BUILD

Teams stitch five products to get what the path gives you natively.

Five contracts, five dashboards, five places the story can disagree — and still nothing that can stop a bad request.

LLM gatewayROUTES, CAN'T GOVERN
Observability toolWATCHES, AFTER THE FACT
Guardrail serviceSCORES, CAN'T BLOCK
FinOps platformSEES THE INVOICE ONLY
Developer analyticsCOUNTS ACTIVITY, NOT ROI
COLLAPSES INTO
Wardin

One gateway in the request path. Policy, ledger, evidence, and outcomes are properties of the same record — they can't disagree.

ONE CONTRACT · ONE DASHBOARD · ONE SOURCE OF TRUTH
05 — ONE RECORD, THREE READERS

The same receipt answers three different questions.

SECURITY & COMPLIANCE ASKS
"What left the building?"

Nothing un-scanned. Every prompt passed PII and injection checks in-path, and the signed chain proves it — per request, mapped to the framework, kept for the full retention window.

FINANCE ASKS
"Who spent it, on what?"

Every token debits a team and a key at request time. Budgets warn at 80% and enforce at 100% — the ledger closes clean without a reconciliation sprint.

LEADERSHIP ASKS
"Is it actually working?"

Spend lines up against PR acceptance and delivery signals — not token volume. You see which teams turn dollars into shipped work, and which just burn.

06 — QUESTIONS, ANSWERED

Before you ask sales.

What is Wardin, in one sentence?

The audit-grade evidence layer for AI: one gateway that enforces policy before a request reaches a provider, meters every dollar to a team and a key, and signs a tamper-evident receipt of what happened. That one receipt answers three separate questions — security asks "what left the building," finance asks "who spent it, on what," leadership asks "is it working" — instead of three stitched-together tools each answering one.

How is Wardin different from LiteLLM or Portkey?

LiteLLM and Portkey are proxies — they route requests. Wardin sits in the request path to enforce: budgets hard-stop with an atomic Redis check before the request reaches a provider, policies (model allowlists, prompt-injection screens) block in-path, and every request emits a signed, tamper-evident receipt. If you only need routing, they're great; if you need audit-grade evidence and finance-grade cost attribution from the same request, that's Wardin.

How is it different from Helicone or Langfuse?

Observability tools tell you what happened after the money is spent — they can't block a request, and they don't produce evidence an auditor would accept. Wardin is enforcement plus evidence: the same gateway that meters cost per team, key, and session can refuse the request that would blow the budget, and signs a hash-chained receipt of the decision either way.

Does using Wardin make us EU AI Act compliant?

No — no software makes an organization compliant; that's a legal determination for your counsel, not a feature we sell. What Wardin produces is audit-grade, Art-12-grade runtime records for gateway-routed traffic: a signed, hash-chained, framework-mapped receipt for every request the gateway handles. That's evidence, not a compliance attestation, and nothing here is described as shipped until it's actually shipped. Wardin also doesn't hold — and won't claim — certifications like SOC 2, ISO 27001, or HIPAA. Our SOC 2 and ISO 27001 evidence packs map signed receipts to those frameworks' controls as input to your own audit — evidence for an examination, never a claim to hold the report.

Is Wardin open source?

Not yet. The gateway may be open-sourced later, but we won't put "open source" on the site until there's a public repo you can actually clone.

Does it work with Claude Code, Cursor, and other agentic tools?

Yes — point ANTHROPIC_BASE_URL (or your tool's base-URL setting) at the gateway with a Wardin virtual key. No client code changes. Agentic traffic is grouped into sessions so you see cost per task, not noise across thousands of tool-call requests.

Which model providers are supported?

Anthropic, OpenAI, AWS Bedrock, and Google Vertex (Gemini), with configurable fallback routing across them. Bring your own provider keys.

How do I get access?

Join the waitlist. We onboard teams in small batches so early support stays personal — reply to the welcome email and tell us what your AI cost chaos looks like, and we'll prioritize you.

Put Wardin in the path. Everything downstream gets simpler.

Point your SDK at one base URL. Policy, ledger, evidence, and outcomes start on the first request.

base_url = "https://gw.wardin.ai/v1"
EARLY ACCESS · NO CREDIT CARD REQUIRED