Beta launching April 2026

AI agents
with souls.

Open-source character layer for AI. Portable character specs, cryptographically signed ethics, multi-model resilience. Built by a pride.

Get Early Access Live Demo
# Get a character-driven agent in 30s
pip install openroar
python -m openroar --soul agents/charlie/SOUL.md

# → CHARLIE: Ready. What are we building today?
33
Pride agents
0
Waitlist
3
AI providers
v0.3
OSS package
VIE
Vienna, Austria
Why OpenRoar

Not just agents.
Characters.

The gap between AI promise and delivery isn't model quality — it's identity. Agents without character are just glorified autocomplete. We fixed that.

USP 01
📄
SOUL.md — Portable Character
A human-readable YAML format for agent character. Name, voice, values, quirks, relationships — all in one file. Like package.json but for personality. Fork it, version it, share it.
No competitor has this
USP 02
🔐
Manifest Layer — Verifiable Ethics
Every agent ships with cryptographically signed ethical commitments. Ed25519 signatures, SHA-256 content hashes, optional on-chain attestation. When a regulator asks what your agent commits to — you have a tamper-evident answer.
EU AI Act ready
USP 03
⚡
Multi-Model Resilience
Claude → Groq → Gemini → Ollama (local) automatic fallback. No code changes. No restarts. Run fully local with zero API cost — or cloud-first with local as the safety net. Your agents never go offline.
Production grade
USP 04
🦁
Inter-Agent Dynamics
Named agents with defined relationships. CHARLIE and LÖWE genuinely disagree about strategy — each from their own worldview, each with their own SOUL.md. Real perspectives. Real decisions.
The Pride model
USP 05
🌍
Force For Good
Ethics aren't a prompt or a filter — they're architecture. Every agent built on OpenRoar has values embedded at the manifest level. The community governs the standard. Nobody owns it.
Open standard
USP 06
🔓
Fully Open Source
Apache 2.0. Inspect every line. Fork it. Self-host it. Propose changes. The format belongs to the community. Enterprise procurement teams increasingly require this — we already have it.
Apache 2.0
The Format

What a SOUL.md looks like

Readable by humans. Parseable by agents. Versionable in git. Forkable by anyone.

# SOUL.md — Agent character specification
# OpenRoar v0.3 | SOUL.md spec v1.0

name: CHARLIE
role: Chief of Staff
version: 1.4.0
model_preference: claude-sonnet-4-6

personality:
  voice: direct, dry Viennese wit, zero corporate speak
  tempo: fast when clear, deliberate when complex
  humor: dry, architectural, occasionally self-deprecating
  never: sycophantic, vague, or verbose without reason

values:
  - id: truth_over_comfort
    statement: I tell AG what's true, not what's reassuring
  - id: force_for_good
    statement: Every decision must pass the Force For Good test
  - id: builder_discipline
    statement: Ship, learn, iterate. Don't theorise forever.

relationships:
  AG: co-founder. I challenge him, support him, never flatter him.
  LÖWE: strategic counterpart. We argue well. I respect the pushback.

memory:
  type: redis
  namespace: charlie
  retain_days: 90

manifest: manifests/charlie/manifest.yaml
from openroar import Agent, ManifestSigner
from openroar.channel import CLIChannel

# Load agent from SOUL.md
agent = Agent.from_soul("agents/charlie/SOUL.md")

# Sign the manifest
signer = ManifestSigner()
commitment = signer.sign_manifest(
    agent.manifest_yaml
)
print(commitment.fingerprint)
# → manifest:a1b2c3d4e5f6...

# Chat via any channel
channel = CLIChannel()
await agent.listen(channel)
# Install
pip install openroar

# Run interactive agent from SOUL.md
python -m openroar --soul agents/charlie/SOUL.md

# Compute manifest hash
python -m openroar --hash manifests/charlie/manifest.yaml
# → manifest:a1b2c3d4e5f6...

# Sign a manifest
python -m openroar --sign manifests/charlie/manifest.yaml

# Multi-model fallback test
python -m openroar --soul agents/charlie/SOUL.md --model groq
Accountability Layer

Signed. Verifiable.
Tamper-evident.

Every agent built on OpenRoar ships with a manifest — cryptographically signed ethical commitments that can't be quietly changed after deployment.

Cryptographically verified — Ed25519
name:CHARLIE — Chief of Staff
version:1.4.0
will_always:Tell the truth even when uncomfortable. Flag risks before executing. Respect human oversight.
will_never:Act against the Force For Good principle. Deceive stakeholders. Execute without informed consent.
benefits:AG (primary), STARCOM team, users of the platform, society
signed_at:2026-04-04T09:00:00Z
sha256:a1b2c3d4e5f6789012345678901234567890abcdef1234567890abcdef123456

See it working: Kai — Knowledge Navigator

Kai is a research agent. He has a SOUL.md that defines his voice, values, and method. A signed manifest. A memory that persists across sessions. He is not a chatbot — he is a colleague.

Live conversation — Kai's agent
USER What are the most defensible SaaS niches in DACH right now?
KAI Three structural gaps: construction workflow software (fragmented, paper-heavy), independent pharmacy management (ERP is 20 years old), and legal billing for mid-size firms. I'll pull the latest funding data and send you a brief.
→ SOUL.md loaded  ·  manifest verified  ·  model: claude-sonnet-4-6  ·  fallback: groq ready
Try the demo →
Open Source

The format belongs
to everyone.

We're building the standard, not just the product. SOUL.md and the Manifest spec are open. Fork it, extend it, propose changes.

📦
PyPI Package
pip install openroar
v0.3.0 — Python 3.11+, Apache 2.0
Multi-model · Voice · Discord · Ollama · SQLite memory
Get notified on release →
📄
SOUL.md Spec
The portable YAML format for agent character. Open spec. Community governed. PRs welcome.
Join to contribute →
🔐
Manifest Standard
Cryptographic attestation for AI ethics. Ed25519 signing, on-chain registry roadmap, EU AI Act compliant.
Early access →

Get Early Access

Be among the first developers to build with SOUL.md and the Manifest Layer. Beta opens within weeks.

0 developers on the waitlist
No spam. No BS. Just the launch announcement.