ryntra-solana-evidence

ryntra-solana-evidence

Enables users to inspect Solana token mint accounts, preflight transfer actions, and verify outcome receipts without holding any keys or signing transactions.

Category
Visit Server

README

Ryntra Solana Evidence Kit

Read what a Solana asset can do to you, preview exactly what a transfer will do, and check a receipt without asking the party that wrote it.

Apache-2.0. Read-only by construction: no key, no signing path, no transaction building, and one RPC method — getAccountInfo. A test enumerates every RPC call and every signing-shaped identifier in the shipped sources on each run, so that sentence is a gate rather than a promise.

npm ci
npm run verify        # lint, typecheck, network-free tests, boundary gate
npm run cli -- inspect 2b1kV6DkPAnxd5ixfnxCpjxmKwqjjaYmCZfHsFu24GXo

That last command reads PayPal USD on mainnet through a public endpoint, with no account anywhere, and prints its eight token extensions with what each one means for somebody holding the asset.

Why

A Token Extensions mint can carry powers a balance does not show you. A permanent delegate lets the issuer move or burn tokens from any account. A transfer hook runs somebody else's program on every transfer. A transfer fee means the recipient gets less than was sent. A default-frozen account state means receiving a token is not the same as being able to move it.

None of that is hidden — it is on chain, in the mint account, decodable by anyone. It is just not read by most of the things people press buttons in.

This kit reads it, says what it means in a sentence a holder understands, and refuses to compress any of it into a score. There is no rating here, and the best verdict the policy engine can return is NO_KNOWN_BLOCKER.

The three tools

flowchart LR
  RPC[("Solana JSON-RPC<br/>getAccountInfo")]
  MINT["mint account bytes"]

  RPC --> MINT
  MINT --> P["<b>inspect_solana_mint</b><br/>Asset Passport<br/>identity · authorities<br/>extensions with provenance"]
  P --> F["<b>preflight_solana_action</b><br/>exact preview<br/>fee math · resolved hook<br/>expected balance deltas"]
  POL["owner policy<br/>13 declared dials"] --> V["policy verdict<br/>13 findings"]
  F --> V
  R[("Outcome Receipt<br/>JSON")] --> VR["<b>verify_solana_receipt</b><br/>schema · integrity<br/>issuer · binding"]

  classDef tool fill:#101418,stroke:#4c8bf5,color:#e8eef6
  classDef data fill:#0d1117,stroke:#3d444d,color:#c9d1d9
  class P,F,V,VR tool
  class RPC,MINT,POL,R data
What it answers Reaches the network
inspect_solana_mint What is this asset, and what can its issuer do to a holder yes, one read
preflight_solana_action What will this exact transfer do, and what does my policy make of it yes, one read
verify_solana_receipt Is this receipt intact, who signed it, and is it about Solana no — entirely local

Each is available three ways: as a typed SDK call, as an MCP tool, and from the CLI.

Layout

lib/solana/          the evidence core — decoding, passport, preflight, policy, receipt
lib/agent-control/   the frozen canonical hashing and the Solana registry adapter
lib/guard/           the one canonical JSON serializer both hashers use
packages/solana-evidence-sdk/   typed calls over the core
packages/solana-evidence-mcp/   local stdio MCP server
examples/solana-evidence-cli/   the runnable CLI
docs/solana/schemas/            published JSON Schemas for non-JS consumers

lib/ is not a copy. These are the same modules ryntra.io runs, extracted with their paths intact — so a result from this kit and a result from the product come from the same code, and an extraction that broke that fails its own tests. lib/stellar/evidence.ts is here for the same reason and no other: the evidence envelope is chain-neutral, and shipping the one implementation beats shipping a copy that drifts. lib/solana/evidence.ts says so where you meet it.

Design rules, and what they cost

Every fact carries how it is known. ONCHAIN_VERIFIED, DECLARED, DERIVED, CONFLICTING, UNKNOWN, UNSUPPORTED. A value with no source is not a value here. The cost is verbosity, and it is worth it: an issuer's own marketing and a ledger read must never wear the same badge.

A sentence describes the setting, not the slot. A mint carrying TransferHook with the all-zero program id has no hook installed — the passport says that, and says who can install one. The generic sentence would have been false on the single most-inspected Token Extensions mint there is.

Amounts are raw base units as strings. A 64-bit amount does not survive a JavaScript number, and a lossy amount is a wrong amount. There is no decimal anywhere on the value path.

An extension nobody wrote semantics for still renders, flagged, with its raw kind. A decoded fact without a pretty sentence is still a fact; hiding it would be the lie.

Nothing throws. Refusals are values with a code and a reason — ACCOUNT_NOT_FOUND, NOT_A_MINT, DECODE_FAILED. "No answer" and "this answer" must not look alike.

Retries cover the transport and never the answer. A response that parses is returned as-is, including "no account", because re-asking does not change facts.

Receipts

A Ryntra Outcome Receipt is an observation of something that already happened — never an instruction, and never proof that Ryntra did it. verify_solana_receipt recomputes both hashes from the receipt's own bytes and checks the Ed25519 signature against the key the receipt carries. Four axes, reported separately:

  • schema — is this the shape of a receipt at all;
  • integrity — do both hashes recompute;
  • issuer — is the signature sound; and, only if you pass trustedPublicKeys, is the key one you trust;
  • binding — does the registry pin name the Solana adapter.

"Signed by someone" is not "signed by Ryntra", and collapsing the two is the mistake that axis exists to prevent. lib/solana/fixtures/receipt-*.json are real artifacts you can run this against, including one that is the signed one with a single character edited.

Trust boundary

  • Non-custodial and keyless. No key is held, read, derived or requested. There is no signing path, and the boundary gate fails the build if one appears.
  • Nothing is executed. No transaction is built or submitted.
  • A verdict is not a safety judgement. It reports the absence of known blockers under a policy the caller declared, and the verdict vocabulary has no word for safety.
  • The endpoint is the operator's choice. No tool or SDK call takes a URL from its caller; provider endpoints come from RYNTRA_SOLANA_RPC_MAINNET and RYNTRA_SOLANA_RPC_DEVNET, and public defaults work with no account.
  • Source-distributed. The root package is private, so an accidental npm publish refuses. No npm package is claimed.

Links

Recommended Servers

playwright-mcp

playwright-mcp

A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.

Official
Featured
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.

Official
Featured
Local
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.

Official
Featured
Local
TypeScript
VeyraX MCP

VeyraX MCP

Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.

Official
Featured
Local
graphlit-mcp-server

graphlit-mcp-server

The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.

Official
Featured
TypeScript
Kagi MCP Server

Kagi MCP Server

An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.

Official
Featured
Python
E2B

E2B

Using MCP to run code via e2b.

Official
Featured
Neon Database

Neon Database

MCP server for interacting with Neon Management API and databases

Official
Featured
Exa Search

Exa Search

A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.

Official
Featured
Qdrant Server

Qdrant Server

This repository is an example of how to create a MCP server for Qdrant, a vector search engine.

Official
Featured