evernode-mcp — Evernode AI Builder

A Model Context Protocol server that lets an AI agent build, check, cost, and deploy HotPocket dApps on Evernode / Xahau. Point any MCP-capable agent (Claude, etc.) at it and ask for a dApp — it scaffolds a deterministic-by-construction contract, heuristically checks it won't break cluster consensus, estimates the EVR lease, ranks live Evernode hosts, generates the deploy commands, and — when the dApp moves value on Xahau — hands the settlement Hook off to the trifecta to be proven safe.

HotPocket is Evernode's contract runtime: it runs your Node.js/WASM contract on every node of a cluster and puts the output + state through consensus. The single biggest beginner mistake is non-deterministic code (Date.now(), Math.random(), fetch(), …) — two nodes compute different results and the ledger stalls. This server is built around catching that.

It is advisory and read-only: it generates files + guidance, never holds keys, never spends EVR, never acquires a lease, never signs or submits, and it connects to the Offledger Cluster Manager rather than replacing its orchestration.

Where it fits — the layer-2 companion to the Hooks trifecta

The Hooks pipeline secures layer-1 Hooks — write → simulate one tx → prove all inputs → watch live:

evernode-mcp is the layer-2 companion: it builds the HotPocket dApps that run on Evernode hosts. Whenever a dApp settles value on Xahau through a Hook-guarded account, it hands off to the trifecta — it never re-asserts settlement safety itself (check_hook_compat / generate_settlement emit the trifecta's prove/install commands, not a safety verdict).

Related MCP server: Pentagonal

Tools

All twelve tools are read-only / advisory. Every tool sets readOnlyHint: true; only the two live OnLedger tools (recommend_hosts, host_diagnostics) set openWorldHint: true (they may reach a live external endpoint, OnLedger). Each tool publishes an output schema, so an agent gets a validated structuredContent shape (not just text) — no guessing field names.

Templates

generate_contract ships 10 templates. Each is deterministic by construction (no wall-clock, no randomness, no network in the contract path; time is the consensus ledger seq ctx.lclSeqNo; state persists only through the contract's own consensused state file) and is dogfooded clean through check_determinism (no HIGH findings) and check_contract_api by the smoke test and the suite.

Each generated file set carries notes including "before deploy: run check_determinism on src/index.js — HIGH findings break consensus."

Why the determinism check matters (and exactly what it does / doesn't catch)

HotPocket consensuses contract output across every node. If two nodes diverge — because you called Date.now(), Math.random(), fetch(), or read process.env, or iterated an unordered collection built in a non-consensused order — consensus breaks and the ledger stalls. check_determinism flags these before you deploy.

It is a heuristic linter, not a prover — a source scan (mostly per-line regex, plus a small cross-line alias pass), so it is guidance, never a guarantee. Design bias (deliberate): for this tool a false-negative (silently missing a real consensus breaker) is the worst outcome, so when a construct could iterate/serialize an unordered collection but can't be proven sorted, it is flagged. A false-positive (flagging safe code) only costs you a justification.

Now covered (each with a why + a concrete fix):

• Wall-clock — Date.now, performance.now, process.hrtime[.bigint], new Date() (HIGH).

• Randomness — Math.random, crypto.randomBytes/randomUUID/randomInt/randomFill[Sync]/getRandomValues (HIGH).

• Network I/O — fetch/axios/got/node-fetch, require('https'|'net'|'dns') (HIGH).

• Per-node env — process.env/process.pid, os.hostname/networkInterfaces/cpus/freemem/loadavg/uptime/userInfo/platform/arch/tmpdir/endianness (HIGH).

• Timers / race — setTimeout/setInterval/setImmediate, Promise.race/any (MEDIUM).

• Filesystem — fs.read*/write*/stat/readdir outside the sanctioned state file (MEDIUM).

• Unordered iteration — for..in, and Object.keys/values/entries + Map/Set for..of (LOW). Sorted views (Object.keys(o).sort(), Object.entries(o).sort(...)) are recognized and not flagged.

• Aliased Map/Set (LOW) — a variable or a member (this.m = new Map(), state.m = new Set()) bound to new Map()/new Set() then iterated / spread / forEach'd / Array.from'd / .entries()/.keys()/.values() on a later line (the cross-line alias pass — now covers member-expression aliases, not just plain identifiers).

• Member-expression for..of (LOW) — for (const x of this.m / state.m / obj.m) even with no Map/Set evidence: order-unprovable, so flagged. Known deterministic array members (user.inputs/outputs) and call expressions (ctx.users.list()) are recognized and not flagged.

• Spread / Array.from materialization (LOW) — [...map], [...Object.values(o)], Array.from(set) that materialize insertion order into an array; suppressed when immediately .sort()-ed.

• .forEach (LOW) — over an Object view or new Map/Set; suppressed when sorted first.

• JSON.stringify of an unordered object (LOW) — a bare object identifier or a spread/merge whose key order isn't provably consensused (the serialized output/state is consensused byte-for-byte). Fixed-key object literals, arrays, primitives, a sorted replacer array, and .sort()-ed arguments are recognized as safe and not flagged.

• Locale / timezone / ICU (MEDIUM) — toLocaleString / toLocaleDateString / toLocaleTimeString, localeCompare, and Intl.*. These depend on the host's locale + ICU collation/format data (and timezone), which differ across nodes — the produced string or sort order diverges. Fix: locale-independent formatting + a code-point comparison (a < b ? -1 : a > b ? 1 : 0), never localeCompare.

• Floating-point literals (LOW) — a non-integer float literal (e.g. 0.1, or a negative-exponent scientific literal 1.5e-3 / 1e-3) or parseFloat( feeding contract math: float rounding / NaN / -0 can differ across engines/hosts. Fix: integer math only (work in drops, Math.floor(a*n/d)). Integer literals, integer-valued positive-exponent literals (1.5e3 = 1500), and integer division/floor are not flagged.

Still out of scope (documented honestly — these are NOT caught):

• Bare float math / untyped division — a / b of two unknown-typed variables (no float literal / parseFloat signal) is too noisy to flag soundly, so it isn't. Only float literals and parseFloat are flagged.

• Deeper data-flow — order divergence behind multi-hop aliases (const n = m), function-return values (const m = makeMap()), Map passed in as a parameter, object spreads merged across several statements, dynamically-built call expressions, or a Map reached through a separate-statement reassignment (let m; m = new Map()). The alias pass covers the direct const x = new Map() and the direct member this.m = new Map() cases, not arbitrary data-flow.

• Known acceptable false-positives — e.g. [...Object.keys(o)].sort() still fires the base iteration-order rule (the spread hides the .sort() from it); a Date.now() used only for a local log; an order-independent reduction over a Map; an honest array iterated as for (const x of this.list) (a member with no array-allowlist entry). These flag safe code (the acceptable direction) — justify or refactor.

So: it catches the breakers beginners hit, and biases toward over-flagging the iteration/serialize classes — it does not prove determinism. Settlement safety is proven separately by the trifecta (xahc-prover). Always test on a real multi-node cluster before mainnet.

Settlement → the trifecta handoff

When a value-moving dApp (escrow / subscription payout / payment_splitter) pays out on Xahau, it does so from the cluster's multisig account. generate_settlement produces the "safe-by-construction → proven-safe" bundle:

1. Cluster-side payout code (xahau/settle.js) — the contract decides amounts deterministically under consensus; signing happens OUTSIDE consensus via the cluster's threshold/multisig signer (no single node holds spend power).

2. The install of the trifecta's already-proven, testnet-validated agent_guardrail Hook on the cluster account, with your per-tx LIM (spend cap, 8-byte big-endian HookParameter) + optional DST (destination lock) — emitted as an unsigned xahc install-tx SetHook to sign offline.

3. The exact xahc prove command to prove the guardrail invariant on your built WASM.

The bundle emits the prove/install commands — it does not assert a safety verdict itself. Even a buggy or compromised signer set cannot exceed the LIM or pay a non-allowed DST: the ledger rejects it. Deploy only on a PROVEN verdict from xahc prove.

Install

Install straight from GitHub — it builds on install (the prepare script runs tsc):

npm install -g github:Hugegreencandle/evernode-mcp

Or clone and build:

git clone https://github.com/Hugegreencandle/evernode-mcp && cd evernode-mcp
npm install        # `prepare` compiles dist/ automatically
npm run smoke      # offline self-test (templates determinism-clean, checker + math work)
npm test           # the full Vitest suite (offline)

Add to an MCP client (e.g. Claude Code / Desktop):

{ "mcpServers": { "evernode": { "command": "evernode-mcp" } } }

Or point it directly at the built entry:

{ "mcpServers": { "evernode": { "command": "node", "args": ["/path/to/evernode-mcp/dist/index.js"] } } }

Usage

Point any MCP-capable agent at the server and just ask, e.g.:

• "Scaffold an escrow HotPocket dApp called vault." → generate_contract

• "Is this contract safe for cluster consensus?" (paste source) → check_determinism

• "Does this contract use the HotPocket API correctly?" (paste source) → check_contract_api

• "Scaffold an oracle dApp that agrees on a price via NPL." → generate_contract (oracle_consumer)

• "Is host rHostAddr… healthy enough to lease?" → host_diagnostics (live)

• "What pattern should I use for a token-gated forum?" → recommend_pattern

• "Find me the 5 cheapest active Evernode hosts in Germany." → recommend_hosts (live)

• "Estimate the EVR to run a 3-node cluster for 720 moments at 2 EVR/moment." → estimate_lease_cost

• "Generate the safe Xahau settlement for my splitter, capped at 50 XAH." → generate_settlement → then run the emitted xahc prove command.

Dev / test / CI

npm run build      # tsc → dist/
npm test           # build + Vitest (offline; mocks the live OnLedger fetch)
npm run smoke      # node dist/index.js --smoke — offline self-test
node dist/index.js # run as a stdio MCP server

• Tests (tests/): determinism (rule coverage incl. the regression floor + new gaps), contractApi (good contract clean + each API-misuse flagged), advisor (lease math, host ranking, error mapping, pattern/deploy branches), templates (per-template build + determinism-clean + per-template invariants), settlement (LIM encoding + trifecta handoff shape), outputSchemas (each handler's real output validates against its published schema), index (end-to-end: every tool driven through an in-memory MCP client, input-schema rejection), hostDiagnostics (healthy / red-flag / not-found / fetch-failure honesty, mocked fetch), and fetch (live-path hardening, mocked).

• CI (.github/workflows/ci.yml): on push + PR to main, runs npm ci, npm run build, npm test, and npm run smoke on Node 20.

• createServer() is exported from src/index.ts so the server can be driven over an in-memory transport in tests without starting the stdio transport.

Honest scope (recap)

• Generates code + guidance; does not acquire leases, sign, or move EVR/XAH. No key custody.

• recommend_hosts fetches live from OnLedger (or ranks a list you supply) — it never fabricates host addresses/specs; on fetch failure it returns empty hosts + a note explaining why, never a fabricated fallback.

• check_determinism is a heuristic source scan — guidance, not a proof; biased to over-flag the iteration/serialize classes; test on a real multi-node cluster before mainnet.

• Settlement safety (Hook spend limits) is delegated to the trifecta, which proves it — this server emits the prove/install commands, never a verdict.

License

MIT © 2026 Dane Brown. Open source; see LICENSE. Not affiliated with Evernode Labs or the Xahau project. check_determinism findings are heuristic guidance, not a guarantee — always test on a multi-node cluster and review before mainnet.