agentleFS
arcium-hqarcium-hq/agent-skills/skills/arcium/SKILL.md

arcium

Skill14 starsChanged 4 months agoMITView on GitHub

Build and debug encrypted Solana applications with Arcium — data stays private during computation, no single party sees it. Use when writing Arcis circuits (#[encrypted], #[instruction]), wiring Anchor programs with init/queue_computation/callback flows, choosing Shared vs Mxe encrypted state, encrypting inputs with @arcium-hq/client (RescueCipher, x25519), or debugging ArgBuilder ordering, nonce, callback, or computation finalization failures. Covers dark pools, sealed-bid auctions, encrypted voting, hidden game state, confidential DeFi, secure randomness, and threshold signing. Also use for getting started with your first Arcium app. SKIP: generic Solana programs without encrypted compute, AES or wallet-only encryption client code, MPC theory not tied to Arcium, or Token-2022 confidential transfers.

We split this skill into the rules it gives an agent

9 rules. About 0 are common, 3 are distinctive, and 6 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 6

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    No MCP?: Use the docs directly: docs.arcium.com/llms.txt for the page index, or llms-full.txt for all docs in one file.

    line 31
  2. Very specific:

    All data must be fixed-size — no Vec, String, HashMap.

    line 59
  3. Distinctive:

    Every computation needs three functions in your Solana program:

    line 88
  4. Very specific:

    NEVER reuse a nonce — every cipher.encrypt() call needs a fresh randomBytes(16)

    line 144
  5. Very specific:

    NEVER combine multiple ciphertexts into one ArgBuilder call — each encrypted scalar is its own [u8.

    line 145
  6. Very specific:

    NEVER omit .x25519_pubkey() for Enc<Shared, T> (silent failure).

    line 146
  7. Very specific:

    Macro string matching: All macro strings must exactly match #[instruction] fn NAME across #[arcium_callback], comp_def_offset(), …

    line 150
  8. Distinctive:

    Prefer arithmetic over comparisons (cheaper in MPC)

    line 161
  9. Distinctive:

    When a computation fails, returns wrong results, or never finalizes — check in this order:

    line 168

≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Arcium

Encrypted computation on Solana via MPC. Data stays encrypted during computation. Your confidential app — a Solana program + Arcis circuits (the Rust circuit framework) + an on-chain metadata account — is an MXE (MPC eXecution Environment). The arcium CLI (wraps Anchor) handles init, build, test, and deploy — use MCP for current flags and options.

Targets Arcium v0.11.x (Anchor 1.0.2, Solana 3.1.10 — unchanged from v0.10.x). Upgrading from v0.10.x? See the v0.10 → v0.11 migration guide — the one app-facing breaking change is queue_computation's new 7th arg callback_cu_limit: u32 (pass 0 for default). v0.11 also adds Arcis enums and Option<T> (additive; not usable as circuit input/output). Coming from v0.9.x? Do the v0.9 → v0.10 migration first (init_comp_def → init_computation_def, @coral-xyz/anchor → @anchor-lang/core, Box<…> on queue-side accounts).

MCP Tools: search_arcium_docs for discovery (returns page path), then query_docs_filesystem_arcium_docs with cat <path>.mdx for full-page reads (e.g., cat /developers/arcis/mental-model.mdx).

No MCP? Use the docs directly: docs.arcium.com/llms.txt for the page index, or llms-full.txt for all docs in one file.

When to Use

Use when:

  • You need trustless computation -- cryptographically guaranteed, no single party sees the data
  • Multiple parties compute on combined data without revealing inputs
  • On-chain state must remain encrypted but computable
  • Privacy: sealed-bid auctions, voting, hidden game state, dark pools, confidential DeFi

Constraints:

  • Fixed loop bounds required (no variable-length iteration)

Mental Model

Arcium apps have three coupled surfaces. Most bugs are mismatches across their boundaries:

SurfaceResponsibilityCommon Boundary Bugs
Circuit (Arcis/Rust)Pure fixed-shape MPC logicVariable loops, dynamic collections, .reveal() inside conditionals
Program (Anchor/Rust)Orchestration: init + queue + callbackMacro name mismatch, callback accounts not writable, wrong ArgBuilder order
Client (TypeScript)Key exchange, encryption, submission, decryptionNonce reuse, missing .x25519_pubkey() for Shared, param order ≠ circuit order

MPC constraints (from how secret sharing works):

  • Both branches of if/else execute unless the condition is a compile-time constant — cost = sum of both branches, not max. Same rule applies to non-constant match arms and if let — all reachable arms execute.
  • Loops must have fixed bounds — no while, break, continue
  • Comparisons are expensive; arithmetic (add/multiply) is nearly free
  • .reveal() and .from_arcis() cannot be called inside conditionals (exception: compile-time constant conditions; also forbidden inside non-constant match arms and guards)
  • All data must be fixed-size — no Vec, String, HashMap; use [T; N]
  • Pattern matching (match, if let, matches!) is supported in v0.10+. Last match arms cannot have guards. Let chains require edition = "2024" in encrypted-ixs/Cargo.toml.
  • Enums and Option<T> are supported in v0.11+ (full Option combinator surface — .unwrap_or, .map, .and_then, etc.). Two constraints: neither can be a circuit input or output, and enum discriminants cannot be set explicitly.

Intent Router

Identify what you're building, then read the linked reference before coding. For API details, CLI flags, deployment, and versions, use MCP directly.

IntentReadMCP Query
First Arcium appminimal-circuit.md"hello world tutorial"
Core concepts / terminology (MXE, Arx, comp def)core-concepts"arcium core concepts"
Choose a pattern (stateless, stateful, multi-party)patterns.md"arcium examples"
Circuit syntax (#[encrypted], #[instruction])patterns.md"arcis encrypted instruction"
Shared vs Mxe encryptionSee Encryption Context below"Shared vs Mxe encryption"
ArgBuilder ordering / ciphertext errorstroubleshooting.md -- ArgBuilder Ordering Errors"ArgBuilder encrypted plaintext"
Callback not firing / computation stucktroubleshooting.md -- Computation Never Finalizes"arcium_callback queue_computation"
Nonce / decryption errorstroubleshooting.md -- Nonce Errors"RescueCipher encrypt nonce"
Client-side encryption (RescueCipher, x25519)minimal-circuit.md -- Test section"RescueCipher encrypt nonce"
Threshold signing / secure randomnesspatterns.md"MXESigningKey sign" or "ArcisRNG"
Re-encryption / sealingpatterns.md"arcium sealing re-encryption"
Deployment (devnet/mainnet)—"arcium deploy cluster-offset"
CI / GitHub Actionssetup-arcium"arcium github actions ci"
Version / installation requirementsinstallation docs"arcium installation anchor solana"
Upgrading to v0.11.x (callback_cu_limit, enums/Option)migration guide"v0.11 migration"
Closing MXE / comp def to reclaim rent—"deactivate close computation definition"

Core Pattern: Three Functions

Every computation needs three functions in your Solana program:

FunctionPurposeWhen Called
init_<name>_comp_defInit computation definition — on-chain account holding the circuit's compiled bytecodeOnce per instruction
<name>Build args + queue computationEach request
<name>_callbackHandle result from Arx nodes (the cluster's MPC compute nodes)After MPC completes
const COMP_DEF_OFFSET_FLIP: u32 = comp_def_offset("flip");

// 1. INIT (once per instruction type) — v0.10+ helper is `init_computation_def` (2 args)
pub fn init_flip_comp_def(ctx: Context<InitFlipCompDef>) -> Result<()> {
    init_computation_def(ctx.accounts, None)
}

// 2. QUEUE (each computation)
pub fn flip(ctx: Context<Flip>, offset: u64, ...) -> Result<()> {
    let args = ArgBuilder::new()...build();
    queue_computation(ctx.accounts, offset, args,
        vec![FlipCallback::callback_ix(offset, &ctx.accounts.mxe_account, &[])?],
        1, 0, 0, // num_callback_txs, cu_price_micro, callback_cu_limit (v0.11: 7th arg, 0 = default)
    )?;
    Ok(())
}

// 3. CALLBACK (after MPC completes)
#[arcium_callback(encrypted_ix = "flip")]
pub fn flip_callback(ctx: Context<FlipCallback>,
    output: SignedComputationOutputs<FlipOutput>) -> Result<()> {
    let result = output.verify_output(...)?;
    // Use result...
}

Encryption size: RescueCipher encrypts any scalar to 32 bytes regardless of type. Formula: ciphertext_size = 32 * number_of_scalar_values. See troubleshooting.md for the full size table.

Encryption Context

Enc<Owner, T> — Owner is who can decrypt. Choose before writing the circuit.

ScenarioUse
User submits input / result sealed back to that same userEnc<Shared, T>
Internal state users shouldn't accessEnc<Mxe, T>
State persisted across computations (re-read by any party)Enc<Mxe, T>
Result revealed to everyone.reveal()

Wrong owner fails quietly: persisted/multi-party state must be Mxe — Shared grants reveal capability to one client's x25519 key, leaking aggregate state to that recipient. If a different client key/nonce is used later, decryption yields garbage (MPC cannot raise a decrypt error without leaking). Shared also seals to one recipient; for a public result use .reveal(), not Shared. (Omitting .x25519_pubkey() on a Shared input is the same silent-failure class — see Gotchas.)

Gotchas

Reference during development to avoid common mistakes.

NEVER:

  • NEVER reuse a nonce — every cipher.encrypt() call needs a fresh randomBytes(16)
  • NEVER combine multiple ciphertexts into one ArgBuilder call — each encrypted scalar is its own [u8; 32] call
  • NEVER omit .x25519_pubkey() for Enc<Shared, T> (silent failure); Enc<Mxe, T> skips it

Critical (silent failures)

  • Box-wrap queue-side Arcium accounts (v0.10+): In #[queue_computation_accounts], follow generated templates and wrap heavy accounts: Box<Account<'info, MXEAccount>>, Box<Account<'info, Cluster>>, Box<Account<'info, ComputationDefinitionAccount>>. Unboxed queue accounts can exceed Anchor/Solana stack limits. Callback-side (#[callback_accounts]) can stay unboxed.
  • Macro string matching: All macro strings must exactly match #[instruction] fn NAME across #[arcium_callback], comp_def_offset(), #[init_computation_definition_accounts], #[queue_computation_accounts], #[callback_accounts]
  • ArgBuilder ordering: Calls must match circuit parameter order left-to-right. For Enc<Shared, T>: .x25519_pubkey() then .plaintext_u128(nonce) then ciphertexts. For Enc<Mxe, T>: .plaintext_u128(nonce) then ciphertexts. Missing .x25519_pubkey() for Shared = silent failure.
  • Division by secret zero: Guard divisors with the safe divisor pattern -- both branches execute in MPC, so the division always runs. See patterns.md — Safe Division.
  • Combined ciphertext arrays: Each encrypted scalar needs a separate [u8; 32] ArgBuilder call — do NOT pass [u8; 64] for a two-scalar type. See troubleshooting.md — Ciphertext Size Mismatch.

Warning (wrong results)

  • Nonce reuse: Same nonce for multiple encryptions = garbled output. Use unique randomBytes(16) per encryption.
  • Callback account writability: Pass extra accounts via CallbackAccount { pubkey, is_writable: true } in callback_ix(..., &[...]). Also mark #[account(mut)] in callback struct. Accounts cannot be created or resized during callbacks.
  • Output struct naming: Circuit fn add_together generates AddTogetherOutput. Single returns use field_0 (a SharedEncryptedStruct<1> or MXEEncryptedStruct<1> with .ciphertexts and .nonce). Tuple returns nest field_0, field_1, etc.

Tips

  • Prefer arithmetic over comparisons (cheaper in MPC)
  • Comparisons/divisions are cheaper with narrower types (u64 vs u128); storage cost is identical

Debug Triage Order

Start here when a computation fails or returns wrong results.

When a computation fails, returns wrong results, or never finalizes — check in this order:

  1. Names match exactly — #[instruction] fn NAME must match across #[arcium_callback(encrypted_ix = "NAME")], comp_def_offset("NAME"), and all account macros
  2. Comp def initialized — init_*_comp_def must be called once before any computation
  3. ArgBuilder param order — calls must match circuit fn parameters left-to-right
  4. Shared params include pubkey — .x25519_pubkey() before .plaintext_u128(nonce) before ciphertexts (missing = silent failure)
  5. Nonce is unique — fresh randomBytes(16) per encryption, same nonce passed to program
  6. Callback registered and writable — callback_ix(...) passed in queue_computation call, accounts set in BOTH CallbackAccount { pubkey, is_writable: true } AND #[account(mut)] in callback struct
  7. Environment correct — cluster offset matches network, MXE public key available (retry with backoff), RPC endpoint reliable

For detailed error solutions: troubleshooting.md

Verification Checklist

Pre-deploy gate. Run through before deploying or submitting a PR.

Circuit:

  • arcium build compiles without errors
  • No break/continue/return/variable-length loops
  • #[instruction] fn names are consistent across all macros

Program:

  • init_*_comp_def called before first computation (once per instruction type)
  • Every circuit fn has init + invoke + callback instructions
  • #[arcium_callback(encrypted_ix = "...")] matches circuit fn name exactly
  • Extra callback accounts passed via CallbackAccount { pubkey, is_writable: true } AND #[account(mut)] in callback struct

Client:

  • Unique nonce per encryption (no reuse across calls)
  • ArgBuilder call order matches circuit fn parameter order left-to-right
  • .x25519_pubkey() included for every Enc<Shared, T> parameter
  • Cluster offset matches deployment environment

Deploy:

  • arcium test passes locally before deploy
  • RPC endpoint is reliable (not default Solana RPC)
  • CI pins setup-arcium and explicit Arcium/Anchor/Solana versions for reproducible toolchain installs

Resources

  • MCP tools (primary for API details, CLI flags, deployment, versions): search_arcium_docs + query_docs_filesystem_arcium_docs — docs.arcium.com/mcp
  • Docs: docs.arcium.com/developers — no-MCP fallback: llms.txt (index), llms-full.txt (full)
  • v0.10 → v0.11 migration: migration guide — callback_cu_limit 7th arg on queue_computation, Arcis enums and Option<T>
  • Account lifecycle / closing: account-lifecycle — deactivate-computation-definition → TTL (180 slots) → close-computation-definition / close-mxe
  • Examples: github.com/arcium-hq/examples
  • CI / GitHub Actions: github.com/arcium-hq/setup-arcium — GitHub Action for CI toolchain installs; set Arcium/Anchor/Solana versions explicitly for v0.11.x instead of relying on defaults
  • TypeScript SDK: @arcium-hq/client + @arcium-hq/reader (subscribe to computation events) — ts.arcium.com/api
  • Patterns: patterns.md — curated circuit patterns (stateless, stateful, multi-party, randomness, packing, threshold signing, sealing)
  • Troubleshooting: troubleshooting.md — hard-to-debug errors
  • Minimal working app: minimal-circuit.md — circuit + program + test

Download now (ZIP)

  • Other versions1 project uses a version of this.

    Word for word · 1

    1. xeroc/skills★ 1100% same

    Edited · 0

    None found.

    Wrote their own version · 0

    None found.

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

Reviews can't be read right now.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.