X
Xaish
About · The thesis

Why we built Xaish.

Xaish is a multilateral SaaS platform for access to specialized AI Safety services, plus 51 AI security and payment tools.

AI security and AI-mediated payments are converging. Whoever runs both ends up owning the most valuable real estate in the agent economy. So we built it.

Three opinions we have

Models should be honest about what they don't know. Every Xaish response carries the caveats the tool requires — the model can't leave them out, because the server checks coverage and patches in anything missing. The output is decision support, not a decision, and we make sure the person reading it can tell.

Agents need typed APIs, not chat. Every tool is a single typed POST with a real schema. The OpenAPI surface and the MCP descriptor are both first-class, so an agent can read the catalog, call a tool, and chain the next one without anyone in the middle translating intent into JSON.

Payment AI is the proving ground, so it's free. The 52 AI payment tools come with Paynnt at no charge because payment is where AI either works or it doesn't — refund triage, fraud preflight, invoice reconciliation, dispute drafting. If the model holds up under money, it holds up everywhere. The 73 security tools are gated behind the $10k/month tier, and that bill is what pays for the engine underneath the rest.

What every tool comes with

Four things show up in every endpoint, every time:

  • Schema-vs-prompt alignment. Every literal value the schema accepts shows up verbatim in the prompt. A 125-module audit script blocks the PR if they ever drift.
  • Three mandatory caveats per response. Each tool has three calibration themes the response has to mention. If the model skips one, the server adds it back.
  • Caps on thin input and critical findings. When the input is too sparse to support a confident answer, the disposition collapses to something safer. Same when a critical finding turns up.
  • Credential redaction in and out. API keys, bearer tokens, full card numbers, bank accounts, and crypto secrets are stripped before they touch a model or a log.
The stack, if you're curious
  • FastAPI and Pydantic v2 on the backend, with Claude doing the model work
  • SQLite by default, Postgres when you set DATABASE_URL
  • Solana devnet for Paynnt — solana-py, solders, SPL Token
  • Fernet for at-rest secret encryption
  • Next.js 15, TypeScript, and Tailwind for everything you're looking at right now