Xahau MCP

xahau-mcp

A Model Context Protocol server for the Xahau network with two firsts: it runs a Hook's real WebAssembly bytecode in a local VM (no xahaud node required), and it runs a Hooks-specific static-analysis / security rule engine over it — both fully offline. Around that core it adds read-only ledger access, a Xahau-aware binary codec, an instruction-count fee estimate, network-reward math, governance helpers, and unsigned-transaction builders.

Xahau is the XRPL fork whose flagship feature is Hooks — small on-ledger WebAssembly smart contracts. There was no MCP for Xahau and no static analyzer for Hooks; this is both.

The trifecta — safe Hooks, end to end

Three open-source tools, one workflow: write → simulate one tx → prove all inputs. xahau-mcp is the simulate stage.

Why it's useful

Point any MCP-capable agent (Claude, etc.) at this server and it can:

• See the future before signing — simulate_transaction is a pre-sign flight simulator: every hook an unsigned transaction would trigger runs as real bytecode against live ledger state, with per-hook accept/rollback, decoded emitted transactions, simulated state writes and labeled static engine preflights. Its sibling what_if is a time machine: replay any real historical transaction — with your modifications — at its original ledger. Verified to reproduce a real claim's emitted GenesisMint payout to the drop (72,251,963 drops), test-locked.

• Run a Hook without deploying it — execute_hook instantiates the real CreateCode WASM in a local VM, supplies the Hook API over a simulated transaction + ledger state, and reports the actual accept/rollback decision, return code/string, state writes, emitted transactions and a call trace. No xahaud node needed.

• Audit a Hook before it's installed — paste the CreateCode WASM (or an on-ledger hook hash) and get SARIF-lite findings: missing accept/rollback exit, unguarded loops (_g), unknown env imports, dangerous HookGrants, over-broad HookOn, and more.

• Decode the cryptic HookOn bitmap in both directions — the 256-bit, inverted, active-low mask (with the active-high SetHook bit) is easy to get wrong; here it's verified and round-trip-tested.

• Read Xahau ledger state — accounts, installed hooks, hook definitions, hook state, transactions (with HookExecutions metadata), ledgers.

• Answer the #1 retail question — reward_status tells any account whether it's opted in to Xahau network rewards (Balance Adjustments), the exact XAH accrued — computed with the genesis reward hook's own formula and live parameters, verified to reproduce a real on-chain payout to the drop — when it can next claim, and whether the claim is overdue (late claiming forfeits yield).

• Diagnose an Evernode host — evernode_host_diagnostics automates the official troubleshooting checklist for Xahau's largest operator group: registration, heartbeat liveness (the actual on-chain active rule), reputation, EVR trustline, lease offers, specs and accumulated rewards, in one read-only call.

• — turns an engine result + hook return strings into a plain-English cause and a concrete fix.

Why this is the most advanced blockchain MCP we know of

Strong claim, so here is the checkable evidence (2026-06-11). To our knowledge no MCP for ANY chain — Ethereum, Solana, Bitcoin, XRPL or otherwise — combines even two of these; the closest comparators are cloud-simulation MCPs (e.g. Tenderly's, which simulates on their hosted infrastructure) and standalone analyzers (e.g. Slither, which is EVM-only and not an MCP):

1. Executes real on-chain contract bytecode in a LOCAL VM — execute_hook runs the actual CreateCode WASM with no node, no cloud, no account. Not an ABI wrapper, not a hosted simulator.
2. Publishes a measured, regression-locked fidelity score against chain ground truth — vm_fidelity_report replays 30 real mainnet hook executions: 30/30 agree (100%), 0 degraded, including the foreign-state-reading hook that dominates live traffic. Those 30 are all accept-direction (live Xahau traffic is heartbeat-dominated), and the metric says so itself — it reports the accept/rollback composition and warns that an accept-only corpus can't distinguish the VM from an always-accept stub. The rollback direction is exercised on real genesis bytecode (governance Invoke → rollback) in tests/regression.test.ts. The corpus, the method and the honest history (25% → 0% → 100%) are in docs/FIDELITY.md. We know of no other blockchain MCP that even attempts this.
3. In-protocol static security analysis — a Hooks-specific rule engine (SARIF-lite findings), calibrated against the network's own genesis hooks.
4. In-protocol differential fuzzing — fuzz_hook maps a contract's accept/reject decision boundary in the local VM.
5. Post-mortems real transactions with real bytecode — hook_execution_postmortem replays what actually fired on chain and compares.
6. Reproduces on-chain economics exactly — reward_status re-implements the genesis reward hook's formula and reproduces a real emitted payout to the drop (verified, test-locked).
7. Decodes live governance end-to-end — governance_state shows every seat, vote, tally and threshold of the Governance Game, live.
8. Operational doctors for the ecosystem's real pain: failed-tx diagnosis with cause+fix, Evernode host health, claim-overdue detection.

Every claim above is reproducible from this repo: the corpus is committed, the tests assert the numbers, and the canonical sources (xahaud genesis hooks, evernode-js-client) are cited in code.

Safety posture

• Read-only toward the network. There is no submit and no sign anywhere in this server.

• No key custody. Builder tools never accept a secret/seed and always return an unsigned transaction plus instructions to sign offline (e.g. with xaman or xrpl-accountlib). They default to testnet.

• Honest fidelity. execute_hook runs the real bytecode against a simulated environment. The VM implements a large slice of the 78-function Hook API — nearly the whole XFL float API (verified against float_one; float_root/float_log are the exceptions — they return NOT_IMPLEMENTED), the slot table + STObject subfield extraction (slot_subfield/sto_subfield, byte-exact against real txns), state, otxn_*/hook_*, util_accid/util_raddr/util_verify/util_sha512h, and more. STObject mutation (sto_emplace/erase/validate), util_keylet (account + hook verified against live ledger indexes; offer/escrow/check/ticket/signers canonical + fail-safe), slot_set + foreign hook state (state_foreign/state_foreign_set) with async pre-resolve (execute_hook resolveKeylets:true fetches the ledger objects AND foreign-state entries the hook reads — iteratively, since one resolved read can expose the next — and re-runs), slot_float/float_sto/float_sto_set (STAmount ⇄ XFL both ways — parse a native or issued/IOU Amount into a float and serialize one back; the issued layout below bit 63 is the XFL layout), and 32-byte state-key padding (short keys are left-zero-padded exactly as on-ledger) are now supported. state_foreign_set records the write but does NOT model the on-chain HookGrant requirement; etxn_details serves a disclosed SYNTHETIC placeholder (listed in syntheticCalls, cannot change the accept/rollback decision). What still can't be faithful is honestly recorded: unverified keylet subtypes, meta_slot, and other un-modelled calls return the real code, are listed in , and mark the run — . The VM models the guard budget ( enforces each guard's declared → ), and reports (state writes commit only on , discarded on ). It is a consensus-faithful replica — it has no fee/fuel metering beyond guards, XFL math truncates rather than round-half-up (so 's round-up flag and last-significant-digit results can differ), value-level math is verified only where tested. Hooks with a loop but no guard are (invalid on-chain), and runs are bounded by a (1M cumulative guard calls / 2s wall clock — labeled as a local VM cap, not a consensus limit); always confirm financial/resource hooks on testnet. is , is (legacy — prefer , whose re-implements exactly and reproduces a real on-chain payout to the drop), is .

Tools

Hook intelligence (offline — the core)

Codec / decode (offline)

Ledger (read-only RPC)

Economics / governance

Amendment intelligence (read-only, no admin node)

Audit / forensics

Unsigned builders (no keys, testnet-default)

Cron (scheduled Hook execution)

Install

New here or non-technical? Start with the plain-English tutorial — what it does + cool things to just ask.

Requires Node.js 20+. The server speaks MCP over stdio; your MCP client launches it.

From npm

npx -y xahau-mcp --help      # run without installing (prints usage and exits)
npm install -g xahau-mcp     # or install the `xahau-mcp` command globally
xahau-mcp --smoke            # health check: local data + one live mainnet read

Run with no arguments, xahau-mcp waits for an MCP client on stdin/stdout — it is not an interactive CLI.

Add it to an MCP client

Claude Code:

claude mcp add xahau -- npx -y xahau-mcp

Claude Desktop (claude_desktop_config.json), or any client that takes an mcpServers block:

{
  "mcpServers": {
    "xahau": { "command": "npx", "args": ["-y", "xahau-mcp"] }
  }
}

If you installed globally, "command": "xahau-mcp" with no args works too. To pin your own RPC nodes, add "env": { "XAHAU_RPC_URLS": "https://your-node.example" } (comma-separated; XAHAU_TEST_RPC_URLS for testnet).

The server is read-only toward the network: it reads ledger state over public JSON-RPC, never signs or submits a transaction, and holds no keys. The offline tools (Hook VM, static analyzer, codec, HookOn) need no network at all.

From GitHub / source

npm install -g github:Hugegreencandle/xahau-mcp   # builds on install via `prepare`

Or clone and build:

git clone https://github.com/Hugegreencandle/xahau-mcp && cd xahau-mcp
npm install        # the `prepare` script compiles dist/ automatically
npm run smoke      # health check + a live mainnet read
npm test           # offline test suite

Older versions (≤ 1.4.0) were published to GitHub Packages as @hugegreencandle/xahau-mcp; that scope needs a GitHub token even for public installs. Use the unscoped npm package above instead.

HTTP shim (browsers, wallets, web tools)

The MCP server speaks stdio — browsers, the Xaman webview, and a public "simulate any hook" web tool can't. The HTTP shim (src/http.ts) exposes the same flight-simulator core over plain HTTP. Read-only; never signs or submits.

PORT=8787 npm run http        # or: PORT=8787 node dist/http.js
curl localhost:8787/health

# or deploy the container (non-root, healthchecked):
docker build --target http -t xahau-mcp-http .
docker run -p 8787:8787 -e MAX_INFLIGHT=2 -e XAHC_HOOK_MEM_MB=128 xahau-mcp-http

On platforms that build the last Dockerfile stage (Railway, etc.), use Dockerfile.http (set RAILWAY_DOCKERFILE_PATH=Dockerfile.http) — it's the http stage as the final stage.

Endpoints

Simulate a not-yet-deployed hook — pass candidateCode (the built wasm hex) on /simulate. It replaces tx.Account's on-ledger hook chain, so a freshly-compiled hook runs against the live ledger + full TSH chain BEFORE you SetHook it. (This is what xahc verify and a public simulate tool call.)

Config (env)

Running it publicly

The shim serves untrusted, user-supplied wasm, so each hook execution runs in a throwaway worker_thread with the timeout + memory cap above — a hanging or allocation-bombing hook is terminated without blocking or crashing the main process (security-audited; an infinite loop / unguarded recursion returns 422, not downtime). For a public deploy:

• Run behind a process supervisor (systemd / pm2 / container restart policy) for defense-in-depth.
• Only set TRUST_PROXY when genuinely behind a proxy that sets X-Forwarded-For; otherwise the rate limiter keys on the socket address (setting it lets clients spoof the header and bypass the limit).
• Per-IP rate limit + a global concurrency cap are on by default — tune via env.

Security

Designed defensively and reviewed (npm audit + a danger-surface pass):

• Read-only & no key custody — no sign/submit anywhere; builder tools never accept a secret and only emit unsigned transactions to sign offline.
• No code-exec surface — no eval/Function, no child_process/shell, no filesystem writes, no dynamic require. RPC fetch only ever hits the fixed endpoints in data/endpoints.json (or your XAHAU_RPC_URLS / XAHAU_TEST_RPC_URLS overrides for mainnet/testnet) — never a URL built from tool input, so no SSRF.
• Untrusted Hook WASM is sandboxed — execute_hook/fuzz_hook run hook bytecode in Node's WebAssembly engine, which has no syscall/fs/network access; a hook can only call the in-memory JS Hook-API shims, with bounds-checked memory reads/writes.
• Untrusted-bytecode hardening: before executing a hook the VM refuses modules with an unguarded loop (more loops than _g guard call-sites), with an opcode-scan that couldn't verify the loops, over 128 KiB of bytecode, or declaring more than 512 memory pages; guarded loops are bounded by a cumulative guard budget + wall-clock cap. So an attacker-supplied hook can't hang or OOM a run. Tool output is data, not instructions (treat it as such, as with any MCP).
• Dependencies: npm audit reports only low-severity advisories transitively under xrpl-accountlib's signing libraries (elliptic/bip32/tiny-secp256k1) — code paths this server never calls (it uses only the binary codec).

How it works

• No heavy deps. Three runtime deps: @modelcontextprotocol/sdk, zod, and xrpl-accountlib (used only for the Xahau-aware binary codec; its signing surface is never called). RPC is plain fetch; the WASM reader is hand-rolled and zero-dep; the VM uses Node's built-in WebAssembly engine to run the bytecode with a JS Hook API shim — no WASM toolchain or native deps.
• Real data, regenerable. data/ is built from a live Xahau node's server_definitions and the canonical Hook API list (Xahau/hooks-rs c/extern.h) via npm run fetch:all. The 78-function Hook API catalog carries per-function hazard metadata that drives the analyzer.
• HookOn semantics are verified against the Xahau docs: 256-bit, bit n = tx type n, inverted/active-low (set = does not fire), with bit 22 (SetHook) active-high.

License

MIT © 2026 Dane Brown. Not affiliated with XRPL Labs or the Xahau project. Analyzer findings are heuristic guidance, not a security guarantee — always test on testnet and review hooks independently before mainnet use.