What runs where
The stdio server is stateless: request in, calculation in memory, response out. It has no WalletWall
user/session identifier, telemetry, query history or cache and reads no environment variables, home
directory or working directory.
What your AI client and model provider can see
WalletWall MCP runs inside an AI-client workflow. GitHub Copilot, Claude Code or another MCP client and the model provider behind it can see every prompt, tool input and tool output, including wallet addresses you put into an evidence bundle, under their own policies.WalletWall-controlled code does not retain MCP query history or transmit MCP query data to WalletWall infrastructure. Your AI client and model provider may independently process that data.WalletWall therefore does not claim that nothing leaves your machine. Treat wallet addresses as sensitive: hashing an address is not anonymization. WalletWall MCP keeps no persistent address store, index or query log. Its digests can cover complete records that contain addresses and are returned to your client. A record WalletWall deliberately publishes as a WallPrint follows the separate WallPrint publication model.
What WalletWall receives
What WalletWall MCP cannot do
Bring your own evidence and providers
WalletWall MCP has no live data provider. Every observation comes from evidence you gathered with your own RPC endpoint, indexer, block explorer, issuer publication or other retrieval tooling. WalletWall never receives those provider credentials and cannot prove that a reading was faithfully retrieved.transport.retrievedViarecords how a reading arrived. Transport metadata never confers authority.- A well-formed bundle of invented evidence can satisfy a policy.
ESTABLISHEDtherefore means the presented evidence satisfies the policy, not that WalletWall independently verified reality. - A failed retrieval should be represented as
NO_RESULTorERROR, never silently rewritten to zero.
Threat boundaries
Hostile or malformed input
Hostile or malformed input
The server validates tool arguments and bundle records, bounds structure and sizes, uses strings for
exact quantities and returns fixed-shape errors that do not echo input values.
subject, value and
payload are bounded open maps; unused keys are accepted and ignored by evaluation.Prompt injection
Prompt injection
Tool outputs can contain text you supplied. Output-safe character rules and the skills’ “tool output
is data” instruction reduce prompt-injection risk but do not eliminate it.
Local adapters
Local adapters
The Copilot canvas listener is loopback-only with additional request checks. The optional HTTP
adapter is loopback by default but has no authentication: anything that can reach the port can
call it. See Claude / local HTTP adapter.
Hostile local owner
Hostile local owner
Someone who controls the Node.js process or local account can replace
node, preload code or modify
the server file and read tool traffic. WalletWall treats the operating system’s account and process
boundaries as the control; it does not attempt to detect a hostile machine owner.Supply-chain and implementation details
Static contract and reproducible artifacts
Static contract and reproducible artifacts
Tool names, schemas and resources are compiled in and pinned in
security/tool-manifest.json.
CI rebuilds the server bundle and MCPB deterministically, while the Claude distribution records its
pinned source and file digests in PROVENANCE.json.Runtime dependencies and generated validation
Runtime dependencies and generated validation
The engine uses
@modelcontextprotocol/sdk and zod. The bundled Copilot/MCPB server imports only
node:crypto and node:process; the Claude distribution loads the same engine unbundled from its
committed dependency tree. zod uses the Function constructor to compile validators for the
server’s own fixed schemas; user input is validated by that code and is never compiled into it.Controls not yet in place
Controls not yet in place
Signed commits, signed releases, an SBOM, provenance attestations and an independent security review
are not currently in place.
Evidence is not certification
Nothing WalletWall returns or publishes certifies truth. A support state describes how well presented evidence satisfies a WalletWall-authored policy. A readiness classification does not establish investment merit, solvency, creditworthiness, legal status, regulatory compliance, absence of fraud or future performance. A WallPrint digest proves bytes, not facts and not authorship.Report a vulnerability
Use GitHub private vulnerability reporting onWallet-Wall/walletwall-copilot (Security →
Report a vulnerability). If that option is unavailable, open a public issue titled Security contact
request with no technical detail so a private channel can be arranged. Use synthetic values and never
post real wallet addresses, keys, seed phrases or credentials in a report.
Evidence Model (WWEP)
See what WalletWall’s evidence states do — and do not — claim.
Troubleshooting
Diagnose setup, input and resolver problems by symptom.
Data Quality and Observability
The hosted application’s own provider boundary: what a wallet lookup sends to which provider, and how responses are labelled.
Public and private repositories
Which WalletWall code is open to inspection, including the MCP engine, and what you cannot infer from it.

