Skip to main content

What is Fornax?

Fornax is an evidence-first agent-integrity system for coding agents (Claude Code, Codex). It answers one question:

What should I believe about what this agent is telling me, given the evidence currently available?

Coding agents narrate their own work — "all tests pass," "I fixed the bug," "the build is green." That narration is sometimes wrong, and it is never independently checked by default. Fornax watches a session in real time, captures immutable evidence (tool calls, exit codes, transcripts), and checks the agent's claims against that evidence.

The five-state verdict

Every claim Fornax checks resolves to exactly one of five states — never a made-up trust score, and never silently collapsed into a binary pass/fail:

VerdictIconMeaning
VERIFIED🛡 ✓Evidence confirms the claim.
CONTRADICTED🛡 ✕Evidence directly contradicts the claim.
UNVERIFIED🛡 ?No relevant evidence was observed this session.
REVIEW🛡 !Evidence is ambiguous or partial — a human should look.
UNAVAILABLE🛡 —The runtime can't expose the evidence this check needs (a missing capability is reported honestly, never guessed around).

How it fits together

Coding agent session
│ (hook / transcript tail)

Adapter (thin, provider-specific) ──▶ translates provider events
│ into a canonical AgentEvent

fornax-daemon (one local process)
├─ immutable evidence store (SQLite/WAL, append-only)
├─ verifiers: Claim + Evidence[] + Capabilities → Finding
└─ localhost API: status line, detail command, dashboard

Everything on this path runs locally. There is no cloud dependency in the critical path — evidence capture, verification, and the status line all work with network access disabled. See Privacy & Redaction for what that guarantee actually covers.

Where to go next

  • New to Fornax? Start with Why Fornax? for the problem it solves, then Quick Start to get a finding in your terminal in a few minutes.
  • Already sold? Jump straight to Installation.
  • Want to see the "aha" moment first? Read the First finding walkthrough — a real, reproduced case of an agent claiming tests passed while they hadn't.

Project status

Fornax is v0.0.1, OSS, local-first. The Rust workspace (fornax-daemon, fornax-cli, the Claude Code and Codex adapters, the store, and the verifier) is implemented and tested; cloud/SaaS sync is opt-in and disabled by default (see Privacy & Redaction). Source: horonomy/fornax-core.