AI tools

An AI assistant reads these docs as Markdown. The oracle, an MCP server, answers an assistant's questions about an institution's record.

These docs as Markdown

Every page is available as plain Markdown, built from the same source as the HTML page.

Each Markdown page starts with the page title and description. Each HTML page also points to its Markdown version with a text/markdown alternate link in its head.

Give an assistant llms.txt when it should find the right page itself, and llms-full.txt when it should read everything at once.

Page actions

Under every page title:

• Copy Markdown copies the page's Markdown to your clipboard.
• Open shows a menu: Open in GitHub, View as Markdown, Open in Scira AI, Open in ChatGPT, Open in Claude and Open in Cursor. The assistant links open with the prompt "Read <page URL>, I want to ask questions about it."

The oracle

The oracle is an MCP server in services/oracle. MCP, the Model Context Protocol, is how an AI assistant calls tools that a server exposes. The oracle exposes two tools about an institution's coverage: get_solvency and get_anomalies. It runs over stdio.

Answers from a live orchestrator Coming soon depend on the SDK's orchestrator calls, which the oracle reads through. The tools, their inputs and their outputs are described below as they work once those calls match the orchestrator.

Run it

From the repository root, build it and point your MCP client at the built server:

pnpm exec turbo run build --filter=@solva/oracle

{
  "mcpServers": {
    "solva-oracle": {
      "command": "node",
      "args": ["/path/to/solva/services/oracle/dist/server.js"],
      "env": {
        "SOLVA_NETWORK": "testnet",
        "SOLVA_ORCHESTRATOR_URL": "http://localhost:8080",
        "SOLVA_CONTRACT_ID": "CAFUK25F56VW442VAOHDKGMPYF2BZ3ERVLVAM4H6JTN2WGZPDUQLRTPN"
      }
    }
  }
}

get_solvency

Returns whether an institution's reserves cover its liabilities, confirmed by a read of its own registry on Stellar.

Input

How it answers. It reads the institution's latest stored proof through the SDK's getLatestProof. It then reads the latest proof from the institution's registry on chain with getOnChainLatestProof and checks that R >= L there. If the on-chain read fails, it reports verifiedOnChain: false instead of an error.

Output. One text item holding this JSON:

{
  "status": "solvent",
  "solvent": true,
  "reservesTotal": "16000000",
  "liabilitiesTotal": "9000000",
  "proofId": "1",
  "verifiedOnChain": true,
  "asOf": 1793685990
}

The field and value names are the tool's API. In these docs the same idea is called coverage: solvent: true means the record shows reserves covering liabilities.

get_anomalies

Returns flags raised by comparing each of an institution's proofs with the one before it. It uses only the public totals and timestamps, so it needs no private data.

Input

Output. One text item holding a JSON array of flags:

[
  {
    "kind": "reserve_drift",
    "severity": "critical",
    "proofId": "2",
    "value": -40.62,
    "threshold": -10,
    "message": "Reserves fell 40.62% from the previous cycle.",
    "detectedAt": 1793772390
  }
]

The thresholds are DEFAULT_THRESHOLDS in packages/shared-types/src/anomaly.ts:

Percent changes are computed on the integer totals with BigInt and truncated to two decimal places.

A flag needs two proofs to compare. Today the oracle passes only the latest proof, so get_anomalies raises no flag yet. Reading the full proof history comes with answers from a live orchestrator. The example above shows the shape a flag takes. A move from solvent (16,000,000) to near-breach (9,500,000) is a 40.62% fall, more than twice the 10% threshold, so it is critical.

Next: Changelog.