Verifiable Agent Runtime · v2.7.0

Cryptographic proof of what your AI agent actually did.

VAR seals every agent action into a tamper-evident evidence chain, hardware-attested inside AWS Nitro Enclaves. Any auditor verifies the bundle independently — no access to your infrastructure required.

Request access Request a demo →

Log seq : 1

msg "agent started — fetching TVL data"

stream a4f2c1e8d903b471…

sig 9c2e4b8f1a0d3c7e…

Compute seq : 2

fn echo

inputs_hash 3a7bd3e2f8c1049a…

stream 8f1e3c9b4d07a52e…

sig 4d7e2a1f9c803b65…

Evidence Sealed seq : 3

state d8e3f1a7c29b504d…

ECR 1.0000

result PASS ✓

The construction

Hardware-sealed from first byte to final signature.

Three primitives, stacked in order. Each one is independently verifiable. Together they produce an evidence bundle that no operator — including us — can forge or alter after the fact.

Chain

Rolling SHA-256 hash chain

Every log entry and computation extends the chain: SHA-256(prev_stream ‖ new_data). Once written, nothing can be removed or reordered without breaking every subsequent hash.

Key

Ed25519 signatures from hardware

Each chain snapshot is signed by a key that lives only inside the Nitro Enclave. Its public counterpart is bound into the NSM attestation document — a verifier can confirm the signature originated inside the measured binary.

Verify

Independent bundle verification

Ship the bundle to any auditor. They run apex_verify.py against it — no access to your enclave, no call to VAR-controlled endpoints. The math either checks out or it doesn't.

Where it matters

Two places where unverifiable AI output becomes a liability.

Financial agents

Agents taking irreversible financial actions

AI agents are executing trades, auto-generating Suspicious Activity Reports, managing treasury — each action irreversible and potentially subject to regulatory review. When a federal auditor asks "prove your AI didn't alter this SAR after generation," there is currently no cryptographic answer. VAR gives you one.

Security workflows

AI-generated findings at scale

Security teams are drowning in AI-generated findings they cannot verify at the rate they arrive. The next scarce resource isn't more output — it's trustworthy attribution. VAR's evidence chain records exactly what process produced a finding, on what inputs, under what constraints.

Start in five minutes

No AWS account required.

Simulation mode activates automatically when /dev/nsm is absent — you get a real hash chain and real Ed25519 signatures, with mock PCR measurements in place of hardware attestation.

# 1. Start the gateway
docker run -p 127.0.0.1:8765:8765 \
  ghcr.io/kennethkabogo/var:latest

# 2. Log an agent action
curl -X POST http://127.0.0.1:8765/log \
  -d '{"msg": "agent started"}'

# 3. Run an attested computation
curl -X POST http://127.0.0.1:8765/compute \
  -d '{"fn": "echo", "inputs": {"value": 42}}'

# 4. Seal and verify the bundle
python3 tools/apex_verify.py --self-test