self-host
Free · Apache 2.0
You run R1 yourself. All features. Bring your own compute and your own inference.
- Full harness (PLAN, EXECUTE, VERIFY, COMMIT)
- STOKE protocol and signed receipts
- Cross-family adversarial review
- Custom gates as WASM modules
- Memory tiers L0 to L3 on local disk
- Use your Claude Max, ChatGPT Plus, or OpenAI / Anthropic API keys
- Community support on GitHub
hosted on heroa
Contact us / mo + metered inference
Heroa manages the runtime. You bring the work. Optional.
- Everything in self-host
- Managed runtime, upgrades, and backups
- Receipt retention with configurable SLAs
- Dashboard for sessions, gates, and memory tiers
- Team workspaces with role-based access
- SSO via OIDC and SAML
- Email support
What you bring
R1 is a coordinator, not a model vendor. Cross-family review needs at least two model families. You provide the keys; the harness sorts the rest.
- Implementer key. Claude Max, Claude Code Pro, or an Anthropic API key.
- Reviewer key. ChatGPT Plus, an OpenAI API key, or any non-Anthropic family the harness supports.
- Optional models. Local models via the OpenAI-compatible adapter; Gemini, Mistral, and others through the family registry.
Enterprise
For procurement-driven deployments: Custom. Add SSO, audit log export, STOKE receipt retention SLAs, and sovereign deployment via Heroa BC.
- SSO across OIDC and SAML providers
- Audit log export to S3, GCS, or Object Lock-equivalent
- Receipt retention 7+ years with cryptographic continuity
- Sovereign deployment in Canadian-resident infrastructure (Heroa BC)
- SOC 2 Type II posture; HIPAA BAA on request
Contact for enterprise pricing.
What is and isn't billed
R1 itself never charges per token. The only money flowing to model vendors is what you spend through your own keys. Hosted-tier billing is base subscription plus the metered inference your sessions consume.