AZMX AI · Research
How sovereign AI actually works.
Long-form posts on the architectural choices behind AZMX — per-call approval as a structural property, hash-chained audit logs, customer-rooted license issuance, BYOK economics, and the design of an agent that asks before it acts.
Architecture
Posts on how AZMX is built.
Per-call approval as a structural property
Most agents have an "ask first" flag. AZMX requires a positive ack on every write or shell call by structure. Why that matters for prompt-injection resistance.
Read → ArchitectureHash-chained audit logs in 200 lines of Rust
Tamper-evident logging without a queue, a database, or a third-party service. The chain math, the file format, the SIEM-export shape.
Read → ArchitectureWhy the trust floor doesn't change between tiers
Free, Pro, Teams, Enterprise — the trust property is invariant. The wrapper around the fleet changes. The architectural reason.
Read → ArchitectureCustomer-rooted license issuance
Your fleet's Ed25519 keypair signs licenses. AZMX cannot revoke or impersonate. The threat model, the key-rotation story, the air-gap operation.
Read →Security
Posts on the trust property.
The secret-path screen — what the agent can never read
Pattern-matched OS-layer refusal of .env, .ssh, credentials, vault.yaml. Why we picked patterns over regex, and how custom patterns are deployed on Pro+.
SecurityDLP secret-egress on outbound prompts
Teams+ scans every outbound prompt for high-confidence secret patterns. False-positive rate, performance budget, override workflow, audit trail.
SecurityProvider allowlist as a fleet pin
Centrally limit which AI providers the agent can construct a client for. The org-policy file format, MDM deployment, the fail-safe behavior.
Economics
Posts on the BYOK shape.
BYOK is the only honest model
The vendor incentive problem with usage-billed AI tools. Why pricing AZMX as a fixed license aligns our interests with yours.
Read → EconomicsThe 75% reduction nobody talks about
Per-seat coding tools blend their inference cost into the seat fee. We unblend it. The math on a 50-engineer team.
Read → EconomicsWhat the next decade of developer-productivity billing looks like
Provider-direct, local-first, capex-shaped. We sketch the curve.
Read →Subscribe to the engineering log.
Posts go up on the blog. Major architectural shifts also publish to the GitHub releases page.