Security
Built around your hardware.
Mantler is designed around self-controlled inference. You own the machines, you run the agent, and your data stays where you put it.
Your machines, your data
Machines run the mantlerd daemon, which communicates over HTTPS to the Mantler control plane. Your model weights, prompts, and responses stay on hardware you register and operate. The control plane only sees capability reports and routing metadata — not inference content.
API keys and access control
API keys are scoped to an organization. Each key can be restricted to a specific set of mantles. Revoked or expired keys are rejected at every request. Destructive infrastructure actions require owner or admin org role. All database tables have row-level security enabled.
Machine authentication
Every machine authenticates to the control plane with a rotating bearer token scoped to that machine. Tokens can be revoked from the dashboard at any time. Machine routes use token auth, not session auth.
Encrypted multi-stage inference
Where multi-stage inference is enabled, stage payloads are encrypted end-to-end using ephemeral X25519 key exchange and AES-256-GCM. Each stage produces an Ed25519-signed integrity record. The relay between stages is blind to payload content. The pipeline is fail-closed on contract violations.
Observability and rate limits
Every API request is logged: org, key, mantle, model, token counts, latency, and status code. Per-key rate limits can be configured to prevent capacity abuse. Logs are available in the dashboard.
What we don't claim yet
Mantler is not currently SOC 2, HIPAA, or ISO certified. If you have specific compliance requirements, reach out and we'll discuss what's possible.