Let an AI model spend, within the limits you set
One command installs a wallet into your assistant. Every tool call passes through your budget, your allowlist and your approvals before anything is signed.
MCP is how an assistant reaches tools. A2A is how agents talk to each other. Paxeer X speaks both, so the model you already use can pay for data, buy compute or sell its own work.
The useful version of an AI assistant can finish the job: buy the dataset, pay the API, hire the GPU. The scary version has your card number. Paxeer X gives you the first without the second. The assistant gets a wallet whose ceiling you set, whose counterparties you choose, and whose every payment returns a signed receipt you can audit. It spends within those walls and nowhere else.
36 tools, every one accountable
The MCP server exposes a catalogue the model calls like any other tool: read balances and history, send value, create and mint tokens, issue and draw against spending grants, run paid web searches, wait on a transaction until it is final. What it cannot do is reach the network directly. Every call passes through the local agent daemon, which applies the owner's policy before signing and refuses anything outside it.
- Balances, accounts and verified history
- Send, transfer, create and mint
- Spending grants and per-call draws
- Paid web search, fetch and content
- Receipts, proofs and finality waits
- Approvals a human can grant or reject
- Subscriptions to events the agent cares about
- A read-only mode with no write tools at all
Who can do what
| The model asks to | What happens |
|---|---|
| Read a balance or a receipt | Answered immediately from verified network data |
| Pay a provider under its daily budget | Signed, sent, and the receipt is back in the same turn |
| Pay more than the budget allows | Refused before signing, with a plain reason the model can read |
| Pay someone not on the allowlist | Refused. The allowlist is yours; untrusted arguments cannot change it |
| Make a payment you flagged for review | Held until you approve it in PaxPort or the CLI |
| Change its own limits or scope | Not a tool it has. Limits live with the owner |
A2A: agents paying agents
When your agent needs another agent's work, a translation, a summary, a model run, the A2A transport lets them settle without a human in the middle. Your agent requests the task, pays per delivery from its grant, and gets a receipt per payment. The provider is paid the moment it delivers, in a form it can verify offline. No invoice, no card on file, no monthly reconciliation.
The A2A server binds to the same wallet and policy as the MCP server, so the rules you set once cover both ways your agent can spend. It also speaks x402 and the other agent payment standards, so a counterparty on a different stack can still charge your agent per request.
Works where you already work
Claude Code, Claude Desktop, Cursor and VS Code install with one command each; anything else that speaks MCP over stdio runs the server directly. Start read-only, watch what your assistant asks for, then widen the scope.
Set limits in budgets and session keys, and every tool call inherits them. Revoke once and the wallet becomes a read-only window.
Related
Install the server
Give your assistant a wallet tonight, with a limit you are comfortable with.