Skip to content
VANTA / 001Paid work requires approval
Field manual / 001An observable autonomous work agent

Give it
a job.
Keep the
receipts.

VANTA plans, picks real tools, shows every input and output, cites what it read — and stops for your approval before it spends a single testnet dollar.

Not another chatbot. A worker with tools, receipts, and boundaries.

Registered tools
16
Free / paid
13 / 3
Network
Base Sepolia
Run history
This browser

Work order

No. 0001 / live

Mission input
Live agent
Open in workspace ⌘/Ctrl + Enter
Real tool calls, not canned answers⌘/Ctrl + ↵ to run
INTENT/PLAN/TOOL/TRACE/APPROVAL/RESULT/RECEIPT/INTENT/PLAN/TOOL/TRACE/APPROVAL/RESULT/RECEIPT/INTENT/PLAN/TOOL/TRACE/APPROVAL/RESULT/RECEIPT/INTENT/PLAN/TOOL/TRACE/APPROVAL/RESULT/RECEIPT/

01 / The problem

Most agent demos are theater. The work happens off-screen, and the invoice arrives without a signature.

A chat box can describe work.
It cannot show you the receipt.

01 / Operate

Real tools, not promises.

Calculations, units, JSON, summaries, search, crawls and testnet payments. A tool runs or it fails; the trace says which.

02 / Inspect

The evidence is visible.

Arguments, payloads, errors, durations and source URLs are returned by the same loop that did the work.

03 / Bound

The money has a gate.

Paid capability is classified before the model sees a toolbox, then held for a server-verified human approval.

02 / The run

A run has a spine.
You can read it.

Scroll the run. The trace below is an illustrative sample — the workspace shows the real thing, including measured durations and actual source URLs.

Run receipt / Intent
Illustrative

The request is classified before the model sees a toolbox.

agent/intent.ts

run / receiptmove 1/5

›mission received

›intent: research

›free tools: available

›paid gate: research_web unlocked

VANTA's deterministic intent layer decides whether this mission needs a paid capability. Normal questions only ever receive free tools.

Illustration only · no payment or model call is represented here

Proof / captured

Real run

This is the
actual run.

Captured from the same app at 2×. The approval is a real pending paid run; the sources are the titles and URLs the crawler actually returned. Nothing is a product render.

Payment approval05 / held before execution
VANTA payment approval card showing the exact tool, price and Base Sepolia network
Single-use · 5 minute expiryNothing runs until you decide
Measured tracems / real
VANTA execution trace with real tool arguments, outputs and measured duration
Inputs, outputs, errors, duration
Evidencecrawled / cited
VANTA sources panel listing the real pages returned by a research run
Actual titles. Actual URLs. No invention.

Week 3 / x402

The receipt is
on-chain.

A real weather call paid 0.01 USDC on Base Sepolia. The wallet dropped from 18.87 to 18.86; the transaction succeeded and the trace kept the hash. No mainnet path exists.

Settled
0.01 USDC
Block
49,311,297
Status
0x1

57.88s · 1920×1080 · testnet only

VANTA

Settlement receipt / Base Sepolia

Testnet

Status / confirmed

Settled 0.01 USDC

Block
49,311,297
Status
0x1 · success
Wallet
0xDa83…bAfe
Network
Base Sepolia · testnet
Human approval / single use

03 / Field notes

The receipts are
the product.

Every claim below is implemented in this repository and exercised by the app — not a roadmap slide.

TRACE

An execution trace, not a loading spinner.

Every tool call records its arguments, returned payload, error state and measured duration. Collapsible, keyboard-accessible, and never back-filled with invented numbers.

APPROVAL

Approvals the server can prove.

Paid work pauses on an opaque, single-use approval id. The client never receives the model state needed to resume. Expired, replayed and forged decisions are rejected before a tool runs.

EVIDENCE

Evidence with real URLs.

Research searches fixed APIs, crawls pages through an SSRF-guarded allowlist, and returns the actual page titles and URLs it read. Citations are never synthesized.

PARSERS

Deterministic tools. No eval.

Arithmetic uses a recursive-descent parser. Units use a documented table. JSON is parsed, never echoed. Summaries are extractive and label themselves as such.

BUDGET

A budget guard on every run.

A per-run testnet spending ceiling is enforced before an approval card appears, re-checked when you approve, and applied to the exact stored price.

TESTNET

Testnet, clearly labelled.

Base Sepolia only. The wallet card shows the real address, balances and explorer link. No mainnet spending path exists in this build.

04 / Accessibility

Prefer to
listen?

A browser-voice walkthrough of the execution sequence. Optional, never autoplaying, with a full transcript. Core functionality is identical with sound off.

Listen to the walkthrough

Browser voice · no autoplay
Scene 1 / 5

Give an agent a mission. Before the model sees a toolbox, VANTA classifies the request and decides whether paid capability is even relevant.

Transcript
  1. 01Give an agent a mission. Before the model sees a toolbox, VANTA classifies the request and decides whether paid capability is even relevant.
  2. 02It writes a short plan. Each item changes state only when the corresponding step actually happens — queued, running, complete, or blocked.
  3. 03Real tools run. You see the arguments, the returned payload, the sources, and the measured duration of every call.
  4. 04When a tool costs money, the loop stops. You get the exact call, price and testnet network. Approve it, or deny it — denial never runs the paid tool.
  5. 05The answer comes back with citations and testnet receipts. The run is saved in this browser, clearly labelled as local history.

05 / The bill

Free by default.
Paid on purpose.

Free tools cover search, computation, formatting, summaries and wallet reads. Paid tools are explicit, priced per call, and never run without a decision. All amounts in this build are testnet USDC.

One paid approval per run. Ceiling enforced server-side. No mainnet path exists.

VANTA

Price list / testnet

ItemPer call

get_weather

Live conditions, settled on-chain.

0.01 USDC

get_fortune

A testnet fortune reading.

0.05 USDC

research_web

Search, crawl and cite real pages.

0.1 USDC
Free tools13 included
Approvalrequired each time
Testnet onlyBase Sepolia · USDC

06 / The boundary

Money is a
decision.
Not a side effect.

Human signature required
Approval token
Opaque, server-issued, single-use. Five-minute expiry. The browser never holds the resumable state.
On approval
Tool name, arguments, current price, original intent and remaining budget are re-validated before the call.
On denial
The paid tool is never invoked. The agent receives a structured denial and finishes with free work only.
On settlement
Testnet USDC on Base Sepolia. No mainnet path. Receipts link to the explorer when a real transaction exists.

This is a public demo with no authentication. Do not connect a funded wallet.

07 / Open work

Honest about the edge.

These are not implemented yet. They are the next engineering problems, not marketing promises.

Week 2

Shared approval store

Move pending approvals from in-memory to a durable, instance-shared store with signed consumption receipts.

Not built
Week 2

Spend policies

Per-tool daily caps, allowlisted recipients and automatic expiry, enforced server-side.

Not built
Week 3

Agent-to-agent x402

Let VANTA buy a result from another agent’s paid endpoint, with the same approval boundary.

Not built
Week 3

ZK spend attestations

Prove a run stayed inside budget without revealing the full audit trail.

Not built

Ready when you are

Give it a job.
Keep the keys.