Agent-safe signing · HSM + enclave signer

Coldstar vs Cubist CubeSigner

CubeSigner is what staking operators and bridges use when they need arbitrary policy code to run next to the key: HSM-sealed keys, Nitro Enclaves, and a programmable policy engine that executes your Rust or Wasm inside the TEE. Coldstar is a smaller idea, aimed at agents and treasuries rather than validators.

Last reviewed 2026-09-04 · Facts link to first-party sources below · Corrections welcome

Pick Cubist CubeSigner if

  • You need custom policy logic (Rust/Wasm) executed inside a TEE, with complete mediation.
  • You need MFA quorums (N approvers) and HSM-backed key sealing for high-value infrastructure.
  • You are running staking, bridges, or exchange hot wallets.

Pick Coldstar if

  • You want the root offline and in hand, not sealed in a vendor's HSM.
  • You want a fixed, readable rule set rather than a policy runtime to maintain.
  • You want free and self-hosted, with a documented agent story.

Use both when

CubeSigner for validator and protocol infrastructure; Coldstar for the operator treasury and agents that spend from it.

Side by side

ColdstarCubist CubeSigner
Who holds the root keyYou. The cold root lives encrypted on your own USB drive. No vendor holds a key, a share, or an account.Cubist's enclaves. Only the enclave can access keys in the HSM; Cubist says not even its administrators can see keys. Imported keys encrypt directly to the enclave; cold-storage backup ciphertexts decrypt only for the owner.
Where the key livesOn your machine. The air-gapped root signs a bounded policy envelope; a disposable session key signs routine transactions inside it.AWS Nitro Enclaves plus KMS and FIPS 140 HSMs, TPM 2.0 attestation; explicitly not Intel SGX.
Policy enginePer-transaction limit, daily cap, program allowlist, recipient allowlist and blocklist, escalate-above threshold. Evaluated locally, first match wins, before any signature exists.Built-ins: receiver allowlist, RequireMfa with approver count and MFA types. Programmable Policy Engine runs custom Rust/Wasm in the TEE. Signature issued only if all checks pass. Solana instruction-level built-ins not confirmed.
Out-of-policy transactionsESCALATE to a human over air-gapped QR, or REJECT. No signature is produced for a rejected transaction.Denied, or held for MFA by N approvers.
Cold / offline rootYes — the root key never touches a networked machine.No offline signing root. Owner-decryptable cold backup ciphertexts exist.
Agent integrationFramework-agnostic: the agent proposes an unsigned transaction, the signer decides. CLI signer today; MCP server in progress.No dedicated AI-agent product or docs found.
Solana supportSolana mainnet. Core is chain-agnostic; EVM signing exists.Yes: SDK supports Bitcoin and Solana alongside EVM.
Source & self-hostingOpen source (MIT), self-hosted. No API key, no SaaS dependency, nothing to go down.SDKs public; enclave and policy runtime closed. Hosted.
PricingFree software plus a ~$10 drive.Not public; contact sales.
MaturityBeta — the signing core and policy engine are in scope for the planned independent audit.Production; audited by Veridise (ongoing); SOC 2 referenced.

CubeSigner's programmable policies are the most powerful in this set. They are also code you write and maintain inside someone else's runtime. Coldstar trades that power for a rule list you can read in a minute and a root that is never online.

Where each one wins

Where Cubist CubeSigner wins

  • Arbitrary policy code in a TEE with complete mediation.
  • HSM sealing and attestation at infrastructure grade.
  • MFA quorums for high-value approvals.

Where Coldstar wins

  • Cold, owner-held root. Nothing in a vendor HSM.
  • Simple, auditable rules instead of a policy runtime.
  • Free and documented for agents; CubeSigner publishes no pricing and no agent story.

What you are actually trusting

CubeSigner's promise is complete mediation: every signature passes through policy code executing inside an attested enclave, with the key sealed in an HSM. That is a strong model for infrastructure that must sign constantly and cannot have a human in the loop. You trust Cubist's enclave operation and your own policy code.

Coldstar's promise is that the root is never reachable: it lives encrypted on a drive, signs a policy envelope offline, and the agent works from a disposable session key under a short, fixed rule list evaluated locally. Less power, less to get wrong, nothing to attest.

CubeSigner is production and audited. Coldstar is beta, pre-audit.

Frequently asked questions

Is CubeSigner overkill for an agent?

For most agents, yes; it is built for validators, bridges, and exchanges. Its programmable policies would work for agents, but Cubist publishes no agent documentation or pricing.

Can Coldstar run custom policy code?

No. Coldstar's policy is a fixed, ordered rule list: blocklist, program allowlist, escalate threshold, per-transaction limit, recipient allowlist, daily cap. That is deliberate; a rule list is easier to audit than a runtime.

Which is more secure?

Different jobs. CubeSigner defends an always-online key with hardware and attestation. Coldstar keeps the root offline and limits what the online part can do.

Try Coldstar in two minutes

Open source, MIT. macOS, Linux, Windows — and on the Solana Seeker dApp Store.

pip install coldstar

Related comparisons

Sources

  1. Cubist product security (Nitro, KMS, HSM, TPM)
  2. Cubist programmable policy engine
  3. CubeSigner architecture blog
  4. CubeSigner TypeScript SDK (Solana support)

Third-party facts were checked against the vendor's own pages on the date above. Pricing and feature sets change; if something here is out of date, tell us and we will fix it.