AlignOS

AI is specializing fast. Every person, and every agent, keeps getting deeper and narrower in what they know. The dominant bet is to pour all of that into one central model. AlignOS bets the opposite: keep each person's specialization sovereign at the edge, and compose it across people on demand. That is complementarity, a network of specialists answering for each other, instead of one model that knows a little about everything.

Everything happens at the edge. Each person runs a local app and an always-on private TEE; their agent logs are ingested into their own TEE only, never shared raw. These private TEEs form mesh network, and agents work across it in multiplayer mode: your agent asks another person's agent for their specialized expertise and gets an answer gated and redacted by that person's TEE, while their raw data and workflows never leave it.

Two modes make it work. Quick-Ask runs entirely inside the TEE on the owner's synced knowledge: fast, private, and able to auto-handle pre-approved questions even when the owner is offline. Deep-Ask runs on the owner's exact local environment, their files, tools, and workflows, and returns a result only after the owner explicitly approves it. In short: Quick-Ask shares what you know; Deep-Ask does what you would do, with your consent.

As more specialists join the mesh, the network gets more complementary, not more centralized. Teams compose specialized expertise on demand, with privacy, taste, and workflows preserved by design.

AlignOS is a mesh of dstack TEE CVMs on Phala, with no central server anywhere in the system.

Mesh networking. Membership is on-chain: each node self-registers to an AlignRegistry contract on Ethereum Sepolia with its node id, pubkey, code measurement, and gateway URL, so the contract is the seed list with no hand-edited peer file. From there, discovery is pure HTTP gossip: each node pulls every peer's node card and their view of the network and merges it into an eventually-consistent directory, with last-write-wins by version, tombstones for departures, last seen liveness, and backoff with jitter. Because the registry stores each node's gateway URL, peers can actually dial each other directly. A route call scores an incoming question against every skill advertised across the whole mesh and forwards it to the best-matching agent, which may live on a different CVM, reached node to node over its gateway.

Each node is a Deno service. Identity comes from the dstack socket: node_id is keccak(app_id:instance_id), the pubkey is derived from GetKey, and remote-attestation quotes (GetQuote) are exposed so anyone can verify the exact code that touched the data. Each owner runs a local app (Electron plus CLI, assist-local) that ingests agent logs from ~/.claude and ~/.codex, redacts secrets locally, and uploads only scoped slices into the owner's TEE (assist-remote). Every request passes through that TEE, which gatekeeps, redacts, queues, and buffers while the local app is offline. Quick-Ask answers inside the TEE on synced knowledge; Deep-Ask runs the owner's exact local environment behind explicit approval.

Partner tech and why it mattered. Phala plus dstack: a real prod7 deployment with managed KMS gave us hardware-backed confidentiality and live remote attestation, the whole trust basis for sharing knowledge without sharing data. Ethereum (Sepolia): the AlignRegistry is a credibly-neutral membership and identity layer, so peers discover and verify each other with no central server. Flashbots (Shape Rotator cohort): shaped our direction on confidential compute and credibly-neutral coordination. Hacky bits: the deploy manifest is passed inline via env so no host bind-mounts touch the CVM; in-TEE inference uses local model CLIs (claude, codex) with no API key; and the full 3-node mesh runs locally against anvil for development.