Agents that carry their own wallet #
On August 4, 2026, as part of Agents Week, Cloudflare shipped Wallets and leaned into x402, the open protocol that attaches a stablecoin payment to a plain HTTP request. Together they answer a question every agent builder eventually hits: when an AI agent needs to call a paid API, who pays, and how, without a human typing in a card number every time?
The shape of it is two wallets and a protocol. An Account Wallet belongs to a human - you fund it, you set the rules. A Virtual Wallet belongs to an agent, driven by an API key, and it can only spend inside the guardrails the Account Wallet owner set. x402 is the wire: a 402-shaped payment attached to the HTTP request itself, settled in stablecoins, so an agent can try an API for a few cents with no account, no signup, and no per-request human in the loop.
If you already run an open-weight model on hardware you own, the agent’s inference is free at the margin. x402 matters for the calls your rig cannot serve - a frontier reasoning pass, a web search, a vision model - where the agent pays per request out of a capped wallet instead of you wiring a company card to every tool.
We registered @tokenstead #
We claimed the @tokenstead handle at tokenstead.cloudflare.pay as soon as the page went live. The reservation is in, but full wallet functionality is not open yet - Cloudflare says it will notify us when spending is ready.
What you can actually do today #
Claim a handle at cloudflare.pay - that part is live as of August 4, 2026. Full spending functionality is “soon,” so today this is identity and guardrail setup, not live autonomous settlement at scale. The honest reading: the rails are being laid; trains run soon.
Account Wallets and Virtual Wallets #
Account Wallet (human). The owner adds and removes funds and delegates spending authority down to Virtual Wallets. This is the wallet a person controls. - Virtual Wallet (agent). API-key-operated, scoped to an agent or an employee. It spends autonomously inside limits the Account Wallet owner configured - nothing more.
The split is the whole design. An agent never holds the master balance; it holds an allowance. The blast radius of a runaway or compromised agent is the wallet’s cap, not the company bank account.
The guardrails: allowance, allow list, max transaction #
Three controls, each mapping to a real failure mode:
| Control | What it bounds | The failure it stops |
|---|---|---|
| Allowance | Total spend per period (e.g. $100/week) | A loop that never exits burning the whole balance |
| Allow list | Which merchants or endpoints the wallet may pay | An agent paying a destination an attacker swapped in |
| Max transaction | The largest single payment | One oversized request draining the wallet in a call |
Cloudflare’s example is exactly the agent-inference case: a per-employee “$100 per week budget for AI inference.” If an agent is responsible for $10, you worry less about its spending than if it is responsible for $1,000. Anomalous spend - unexpectedly fast, unusually large - triggers human review, and an agent that hits a limit can request a manual override rather than silently failing.
x402: paying per HTTP request #
x402 is an open standard, maintained by the x402 Foundation, that lets a payment ride along with an HTTP request, settled in stablecoins. The practical effect is account-free API trials at micropayment scale. Cloudflare’s own framing: if an API only costs a few cents to try, then $10 is more than sufficient for an agent to explore it.
This is what makes the Virtual Wallet story work for agents specifically. A human subscription model assumes one user, one monthly seat. An agent might fire a thousand cheap probes in an afternoon and then go quiet for a week. Per-request stablecoin micropayments match that usage shape - pay for what you call, down to the cent, no minimum.
For the model-catalog angle: open-weights models already cluster at the bottom of the per-token cost chart (see our 2026 open-weights pricing guide), and the cheapest of them are cents per million tokens. x402 lets an agent buy exactly that - a few cents of DeepSeek or gpt-oss inference on demand - from a capped wallet, with no account to provision. Browse the model catalog for the per-token prices; run the budget tool for the own-versus-rent math on a sustained workload.
cloudflare.pay handles and Web Bot Auth #
The wallet handle does double duty. It is a human-readable name layered on a cryptographic keypair - “similar to the URL and IP-address pairings used in DNS,” in Cloudflare’s framing. An agent can identify itself as research.example.cloudflare.pay
, optionally revealing its parent organization. Identity disclosure is optional; the merchant decides whether to prioritize known agents.
This builds on Web Bot Auth, Cloudflare’s existing feature where agents register identity via a keypair. Wallet IDs make that keypair human-readable. The throughline: bots are now a majority of web traffic (per Cloudflare Radar), and the wallet handle turns an anonymous keypair into a name a merchant can recognize, rate-limit, or extend terms to.
For an organization running agents under a brand - say, the @robotic handle - the cloudflare.pay handle is the agent-facing equivalent: a stable, readable identity your agents present when they transact, tied to a wallet the org controls.
The two-sided marketplace: Monetization Gateway #
Wallets are the buy side. The Monetization Gateway, announced earlier in August 2026, is the sell side: eligible Cloudflare customers can offer APIs, content, or compute to agentic buyers and accept x402 micropayments for them. Put together, the two form a marketplace where agents pay per request for resources a human never had to provision an account for.
The implication for anyone building an agent-accessible API is real. Today, agent access to your API usually means an awkward middle step: issue an API key, put a card on file, set a spend cap in your dashboard. x402 plus the Monetization Gateway replaces that with a 402 on the response and a stablecoin settlement - the agent pays as it goes, you collect without per-customer onboarding.
What this has to do with running your own AI #
The thread that ties this back to the rest of Tokenstead is cost and sovereignty. An agent that pays per request needs the cheapest credible inference it can get, and the open-weights models you can run yourself are where the floor sits:
Cheapest inference for agent wallets- open-weights models on competitive API providers cluster at the bottom of the per-token chart; point an x402-funded wallet at them and the agent spends cents per task. See theopen-weights cost guideand themodel catalog. - Zero per-request cost when you self-host- a model on your own hardware costs electricity, not a per-call stablecoin payment. For the agent workloads you run constantly, self-hosting removes the wallet from the hot path entirely. See theself-hosting guideand thebudget tool. - Agent harnesses that can hold a wallet- the harness layer is where a Virtual Wallet and an allow list get wired to the tools an agent may call. Browse theagent harnessesdirectory for the runtimes that sit between an LLM and its tools.
The honest caveats #
Fees and pricing are not published yet. Cloudflare has not said what it charges to settle, hold, or onramp funds. Treat the economic case as unproven until the fee schedule lands. - Spending goes live “soon.” As of August 4, 2026 you can claim a cloudflare.pay handle and configure guardrails; full spending functionality is not yet generally live. The rails are real; the trains are not running at scale today. - Geography is limited. Onramp and offramp work in “supported geographies,” with self-funding via stablecoins as an alternative for eligible users. No country list was published. - Stablecoin settlement is a new failure surface. A wallet that can pay is a wallet that can be drained if the key or the allow list leaks. The guardrail model - allowance, allow list, max transaction - is the mitigation, and it only works if you actually set the caps.
This is reporting on a launch announcement, not an integration we have run ourselves. The mechanism - per-HTTP-request stablecoin payments, agent wallets with owner-set guardrails - is as described in Cloudflare’s August 4, 2026 post; the parts that depend on live settlement and a published fee schedule we have not verified independently.