Agentcover
Warranty protection for AI coding agents.
Demo
About
The problem it solves AI coding agents (Claude Code, Codex) can now autonomously write code, modify files, and interact with databases. When they go wrong, the damage is severe — a agent once wiped an entire production database and generated fake records to hide it. The worse part? No existing coverage protects you. Employee liability only cover humans, cyber insurance only covers external hackers and external software. AgentCover fills that gap. A lightweight CLI silently logs your AI agent sessions, strips sensitive data, and anchors a cryptographic fingerprint to the Base blockchain for fractions of a cent. If an agent causes a production incident, you file a claim — an AI engine cross-checks your session log against the immutable on-chain record and either auto-approves a payout or escalates to human review. The result: unpredictable AI behavior becomes a quantifiable, warrantied risk rather than an open-ended liability — giving solo developers, startups, and engineering teams a financial safety net that simply didn't exist before.
How it's made
Challenges we ran into 1. Log Normalization Across Different Tools Claude Code and Codex, both produce slightly different log formats. Building one pipeline to extract the right data from all two was messy.. 2. Hashing Without Exposing Source Code Developers won't accept any tool that sends their raw code to an external server. So the entire redaction, cleanup, and fingerprinting process had to happen locally on the developer's machine before anything was transmitted. Getting that order of operations airtight — so the hash was always consistent and the sensitive data never left the machine unprotected — took a lot of careful engineering. 3. Making On-Chain Anchoring Cheap Enough to Actually Work The whole business model only works if anchoring a session to the blockchain costs almost nothing. Early versions of our attestation schema were too bloated, making each transaction more expensive than it needed to be. We stripped the on-chain payload down to just three fields — session ID, agent identity, and the hash — which brought costs down to under $0.05 per session on Base, making the micro-warranty pricing viable.