Run. Verify. Anchor.
Deterministic, hash-anchored fintech tools that run entirely in your browser. Think of them as standard candles for agents: reference points you can re-verify anytime. Free and open source.
A2A payments, ISO 20022, AML/KYC, DORA, EU AI Act, Basel III/IV, carbon compliance, post-quantum cryptography. If the rule exists, there's a tool for it.
Browse tools ↗Multi-step OpenChainGraph pipelines where each node's output feeds the next, hash-anchored under one open standard with a typed schema and a published spec. Serverless end to end: chains run in your browser, the endpoint on edge workers.
Read the white paper ↗Every output carries a SHA-256 execution hash, and 693 of 693 deterministic nodes carry a real risc0 groth16 zero-knowledge proof. Recompute any result yourself, offline.
Check a proof ↗The full server-side catalog is callable over MCP from Claude, ChatGPT, or any client, no authentication required. Every call comes back with a receipt.
Docs ↗Paste an agent decision receipt and read its lineage: which model produced the output and whether that is attested or asserted, the agent's autonomy tier, the intent to outcome chain, and the hash that covers all of it.
Render the card ↗Commit-reveal joint randomness with 3 to 7 parties, a paste mode for two real devices, and the last-revealer attack demonstrated in-page. Teaching model, zero network.
Run the ceremony ↗Guiding your agents and chatbots.
Stop paying for hallucinated answers: our MCP and WebMCP tooling return verified results, each sealed with a ZK-Proof. A quick example:
| By hand | Spreadsheet | Vendor black box | Web chatbot | AINumbers | |
|---|---|---|---|---|---|
| Repeat the same calculation | ✗fatigue, keying errors | ✗formula drift, broken refs | ~vendor updates change results | ✗might write a poem about it | ✓ every time |
| Your counterparty re-runs it | ✗they redo it by hand | ~only if they trust your file | ✗no access to the model | ✗their model, their answer | ✓ same hash |
| Inputs stay private during the run | ✓ | ✓ | ✗your data on their servers | ✗your prompt is the input | ✓ all inputs and outputs stay local |
| Works offline | ✓ | ✓ | ✗ | ✗ | ✓ serverless |
| Nothing retained afterwards | ✓ | ✓ | ✗your data on their servers | ✗the conversation | ✓ nothing |
| Receipt with execution hash | ✗ | ✗ | ✗ | ✗ | ✓ §4, every run |
| Underwriting-grade evidence | ✗nothing machine-checkable | ✗no execution evidence | ✗black-box outputs | ✗sampled prose | ✓ receipts underwrite agent work |
| Agent-callable (MCP + WebMCP) | ✗ | ✗ | ~APIs with keys and fees | ~APIs with keys and fees | ✓ no auth |
One endpoint
No signup, no API keys, no billing. Point any MCP client at the AINumbers endpoint and the full server-side catalog becomes callable in a single line of config.
{
"mcpServers": {
"ainumbers": {
"url": "https://mcp.ainumbers.co/mcp"
}
}
}
# or from the CLI
$ claude mcp add ainumbers https://mcp.ainumbers.co/mcp
MCP activity
Questions the workflows answer.
Each workflow is a fixed pipeline of deterministic steps, and every run ends the same way: a §4 Policy Mandate receipt carrying the execution hash. Six examples:
Can this AI agent pay, and under whose authority?
Are we ready for DORA's incident rules?
Is our settlement flow CSDR-safe for T+1?
Would our credit book pass IFRS 9?
What do we owe under CBAM?
Can we operate as a MiCA CASP?
Verify before you trust.
Every result ships with its receipt.