PROVENANCE • AUDIT • TRUST

Cryptographic identity
and provenance for AI agents.

Every action your agent takes is hashed, signed, and hash-chained. An immutable, verifiable audit trail that proves exactly what happened — and what didn't.

Local-first • Ed25519 signatures • SQLite • Works with Grok, Cursor & OpenClaw

Why Abavus?

Today's agents operate on social trust: you believe them because they sound reasonable. That doesn't scale for audits, compliance, debugging production agents, or multi-agent systems.

Abavus makes trust cryptographic.

  • Prove what an agent did with signed, hash-chained logs
  • Search complete interaction history (full-text + semantic)
  • Verify the integrity of the entire chain at any time
  • Query by action, session, time, tool, or content
  • Import existing OpenClaw sessions; auto-log from Grok hooks

Features

🔐

Ed25519 Signatures

Every entry is cryptographically signed by your agent's identity. Tamper detection is built-in.

⛓️

Hash Chain

Each record links to the previous via SHA-256. Any rewrite breaks the chain and is immediately detectable.

🔍

Powerful Search

Full-text search across everything. Optional semantic search with Ollama embeddings for natural language queries.

📊

Session Intelligence

Automatic session timelines, tool usage stats, token counts, cost tracking, and activity summaries.

💾

SQLite Storage

Fast local queries, portable single-file DB (~/.abavus/chronicle.db). Export to JSONL anytime.

🪝

Zero-friction Logging

Install Grok/Cursor hooks once. Every tool call, user message, and LLM turn is captured automatically.

Human-readable dashboard

The local web viewer turns raw cryptographic logs into an explorable audit experience:

  • Overview stats + pending spool
  • Rich session list with turns, tokens, cost
  • Expandable timeline with thinking, tool arguments & results
  • Global full-text search across your entire history
  • One-click export and copyable IDs
  • Chain verification at a glance

Run npm run viewer → opens at http://127.0.0.1:3847

Abavus Chronicle • 19.5k entries • 41.5 MB
19,531
entries
9.1k
llm.turn
7.4k
tool.call
💬 User: Read HEARTBEAT.md…
🧠 Assistant (opus-4-5) · 17k tok · thinking
🔧 read · path: HEARTBEAT.md
🔧 exec · command: …
✓ Chain valid • 19,531 entries

How it works

1
Create identity
abavus init
2
Install hooks
abavus hook:install-grok
(restart your agent)
3
Work as normal
Every tool call, prompt, and response is spooled then flushed into the signed chronicle.
4
Inspect & verify
abavus sessions, abavus session <id>, abavus verify, or open the dashboard.

Rich Personas

Create distinct AI personalities with their own cryptographic identity, strengths, knowledge, and fully verifiable history. Fork them while preserving lineage through snapshots.

Create

abavus create persona researcher
--strengths "analysis,fact-checking"

Manage & Inspect

abavus personas
abavus persona researcher

Fork with provenance

abavus fork persona researcher as senior-researcher

All actions are signed under the persona's identity. Perfect for specialized, long-running agents with auditable expertise.

Quick start

git clone https://github.com/toschdev/abavus.git
cd abavus
npm install

# Create your agent identity
node cli/abavus.js init

# Auto-log every Grok / Cursor action
node cli/abavus.js hook:install-grok

# Import historical OpenClaw data (optional)
node cli/abavus.js import

# Explore
node cli/abavus.js stats
node cli/abavus.js sessions
node cli/abavus.js session <id-prefix>

# Beautiful local dashboard
npm run viewer
# → http://127.0.0.1:3847

Data lives in ~/.abavus/ (keys + chronicle.db). Everything is yours.

CLI highlights

Personas

create persona X
personaspersona X
fork persona X as Y

Identity & Chronicle

init id
stats verify search

Sessions & Audit

sessions session <id>
abavus session or viewer

Advanced

ask "query" (semantic)
embedwatchrate

Ready to make your agents accountable?

Clone the repo, create an identity, install the hooks, and never wonder “what did it actually do?” again.

View on GitHub
AGPL-3.0 licensed • Copyright (C) 2026 Tobias Schwarz (toschdev) • abavus.ai