Yang Yang Xie
a92bbcd782
The README was written when Aegis was a monitoring and gateway tool, and it has been describing that product for three months while the thing itself grew into a platform. The headline still read "Self-hosted Envoy gateway · AI threat analysis · TLS automation · Real-time dashboard" — all true, all now one subsystem of several, and none of it the reason someone would choose this. README now leads with what it is: one control plane from the bare metal up, air-gapped by default, with the customer's model weights staying on the customer's storage. The gateway keeps its place; it just stops being the whole story. A note records the scope change rather than pretending the old framing never existed. ROADMAP is rewritten, not amended. It listed a container manager, a Docker registry, an embedded DNS server and AI-driven deployment orchestration as future phases — all four shipped. A roadmap that describes delivered features as upcoming undersells the product to exactly the reader who bothered to open it. It is now built / next / then, matching the private status doc so the two cannot drift. CHANGELOG gains 1.3, covering three months: bare-metal provisioning, the Kubernetes lifecycle with air-gapped installs validated live, Depot with pluggable storage backends and reference-only artifacts, GPU-aware workloads, Owl's risk-classified tool surface, and the auth middleware rewrite that replaced a method-blind public-route match. FIXED A BROKEN QUICKSTART, which is the part that mattered most. The published docker-compose.yml exposes only 8765, while agents dial in on 8766 and the overlay needs 3478/udp and 51820/udp. Anyone following the new one-line enrolment instructions would have watched the installer succeed and the node never appear. Both ports are now published with comments explaining when they are needed, and AGENT_HOST is present with a note that it must be set to something the nodes can actually resolve before enrolling any. Every internal link in README and ROADMAP verified to resolve; compose file validated. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01B9dzsPuPpFbdgCuUDgsG8L
4.3 KiB
Aegis Roadmap
Updated: 2026-08-20
Three sections: what exists, what makes it sellable, what makes it defensible. Dates are deliberately absent below the first section — a roadmap with dates on unstarted work is a wish list.
The previous version of this file listed a container manager, a Docker registry, an embedded DNS server and AI-driven deployment orchestration as future phases. All four shipped. It is rewritten rather than amended, because a roadmap that describes shipped features as upcoming is worse than no roadmap.
Built
Available today and running on a live nine-node fleet.
Fleet — one-line enrolment, per-node mTLS identity from the hub's own CA, digest-verified agent upgrades with systemd auto-rollback, idempotent repair, activity log.
Networking — WireGuard overlay mesh, STUN NAT traversal, automatic LAN-direct routing with mesh fallback, embedded fleet DNS covering hosts and bridge containers.
Gateway — Envoy with a full xDS control plane, SDS certificate management, automated TLS issuance, staged pending→apply configuration, rate limiting, maintenance pages, one-click Expose.
Bare metal — Redfish/BMC power control, UEFI HTTP Boot with no L2 adjacency required, serial console, libvirt/KVM provisioning with cloud-init.
Kubernetes — create, scale, upgrade, back up, restore and tear down via a reconciler, on k3s or kubeadm. Air-gapped installs validated live on a 5-node HA cluster.
Models and data — catalogue with versions, digests, lineage and per-owner quotas; HuggingFace and URL ingest; LAN staging so the hub egresses once; pluggable storage backends (local, mounted share, cloud bucket) switchable at runtime; reference-only artifacts that catalogue without copying.
Workloads — GPU-aware scheduling, managed volumes, notebooks, serving, promote-to-catalogue.
AI operations — Owl acting through risk-classified tools with prompt-injection defences; AI threat analysis over gateway access logs.
Next — what makes it sellable
The gap between "impressive" and "a customer can run this without us".
Multi-user, roles and service accounts. Today there is one operator account
and authenticated means admin everywhere. Ownership already threads through
the data model — artifacts, workloads and projects all carry an owner, and
quotas are per-owner — so this is adding an identity layer on top of existing
scoping rather than retrofitting ownership. This is the single largest gap.
Training as a managed workload. The GPU scheduler, artifact capture, managed
volumes and result promotion all exist; training is simply not modelled as a
workload yet, so a run is still started by hand and observed with tail -f. The
defining workload of an AI platform should not be the one the platform cannot
see.
Control-plane backup and documented recovery. Cluster backup and restore exist. The hub's own state does not have an equivalent a customer could follow.
Audit trail. Node lifecycle is recorded; a general who-did-what across Depot, clusters, gateway and secrets is not.
A quickstart a stranger can complete. If we have to install it, it is a service rather than a product.
Then — what makes it defensible
Control-plane high availability. One hub and one database today. Workloads survive a hub outage; management does not.
Serving as a product surface — autoscaling, canary rollouts, token metering. Serving works; it is not yet something you sell by the seat.
Experiment tracking and lineage across runs, so results are comparable rather than merely stored.
Supply chain — signed artifacts, SBOM, dependency CVE scanning.
Scale validation well beyond a single site. Nine nodes is what we have tested, and it is the number we quote.
Deliberately not doing yet
- Multi-region / federation. One site per hub until one site is excellent.
- A custom scheduler beyond GPU awareness. Kubernetes exists; we orchestrate it rather than replace it.
- Billing and metering. Not until someone is billed.
- A Windows agent. Linux edge is the market.
Questions about any of this are welcome at aegis.jerxie.com, where an assistant answers from this documentation — including the parts about what is missing.