AI agents act without proof. They call tools, get results, and nobody can verify what actually happened. Tressellate fixes this — a layered framework that gives agents structured access to Hiero so every action is cryptographically auditable.
The LLM proposes. The ledger disposes.
git clone https://github.com/tressellate/tressellate.git
click to copy
An agent calls a tool, gets a result, and has no way to verify the response wasn't spoofed, manipulated, or fabricated. That's fine for weather data. It's unacceptable for finance, procurement, or carbon verification.
The agent calls a procurement API. It gets back "approved." Was it actually approved? By whom? When? Against which policy? The agent can't know. The auditor can't know. The response is just a string.
The same action returns a Hiero transaction ID with a consensus timestamp and cryptographic proof. The approval, the approver, the policy version — all on-chain. Verifiable by anyone, retroactively, forever.
Infrastructure separated from domain logic. Developers work in their domain language — certificates, milestones, provenance — not the ledger's.
Like tiles in a tessellation, these 7 primitives fit together with no gaps. An AI agent learns the shapes once and reasons about any regulated domain — crop certificates and drug lot certifications are both Certificate tiles.
The LLM's internal reasoning will always be opaque — and that's irrelevant. The model has no unmediated authority. It can only call tools with defined parameters, and every action is validated by deterministic layers below.
Every agent–tool interaction uses explicit JSON-RPC schemas with defined parameters, constraints, and predictable responses. Developers and auditors inspect exact interfaces without touching model weights.
Every MCP interaction is inherently loggable: tool discovery, calls, parameters, results, errors. A machine-readable chain of actions far easier to audit than opaque neural activations.
Other frameworks let agents blindly trust tool responses. Tressellate grounds every high-stakes action in on-ledger proofs. Even if the AI hallucinates, the outcome is verifiable against the immutable ledger.
Over autonomous iterations, AI agents gradually diverge from intended parameters. Fixed-contract interfaces and per-action ledger attestation make drift detectable by design, not in retrospect.
Each domain implements Layer 4 + Layer 5 — different business logic, same universal primitives. The tessellation in action.
Independent commercial applications being rebuilt on the Tressellate framework — proving the architecture generalizes across regulated industries.
Production deployments requiring advanced resilience and physics-bounded intelligence.
Clone, build, connect your MCP client. The agent discovers tools automatically.
Whether you're building a new domain, evaluating AI governance for your organization, or contributing to the open-source core.