mirror of
https://github.com/axieyangb/aegis.git
synced 2026-09-09 17:25:39 +00:00
docs: rewrite for the product this actually became
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
This commit is contained in:
Yang Yang Xie
@@ -2,6 +2,90 @@
|
||||
|
||||
All notable changes to Aegis are documented here.
|
||||
|
||||
## [1.3] — 2026-08-20
|
||||
|
||||
The release where Aegis stopped being a gateway with extras and became a
|
||||
platform. Three months of work; the highlights only.
|
||||
|
||||
### Fleet & provisioning
|
||||
|
||||
- **Bare metal, end to end** — Redfish/BMC power control, UEFI HTTP Boot with no
|
||||
L2 adjacency required, serial console, and libvirt/KVM provisioning with
|
||||
cloud-init. Power on an unracked server and finish with it in a cluster.
|
||||
- **BMC discovery that discriminates** — a sweep no longer lists every host with
|
||||
443 open. Candidates are labelled with the fleet member already at that
|
||||
address, and proven negatives fold out of the way.
|
||||
- **Node join tells the truth** — enrolment is recorded as an event, rejected
|
||||
agents are reported with the reason and the fix instead of failing silently
|
||||
every 30 seconds, and re-running the installer archives the old identity
|
||||
rather than fighting it.
|
||||
- **Idempotent repair** — `curl -fsSL https://<hub>/reinstall.sh | sudo bash` is
|
||||
safe to run anywhere: it compares digests and does nothing if the node is
|
||||
already current. Agent upgrades verify the download and arm a systemd
|
||||
auto-rollback before replacing the binary.
|
||||
- **Fleet activity log** with per-row prune, and location badges that classify
|
||||
every node as LAN, cloud, mesh, remote or pending.
|
||||
|
||||
### Kubernetes
|
||||
|
||||
- **Full lifecycle via a reconciler** — create, scale, upgrade, back up, restore
|
||||
and tear down, on k3s or kubeadm.
|
||||
- **Air-gapped installs validated live** — a 5-node HA cluster brought up with no
|
||||
egress at all.
|
||||
- In-browser kubectl console and one-click Expose.
|
||||
|
||||
### Models & data (Depot)
|
||||
|
||||
- **Catalogue** for models and datasets with versions, digests, lineage and
|
||||
per-owner quotas; HuggingFace and URL pull-through ingest with progress and
|
||||
gated-model secrets.
|
||||
- **LAN staging** — the hub egresses once and nodes pull from the hub, so a node
|
||||
with no internet access can still receive a 57.9 GB model.
|
||||
- **Pluggable storage backends** — keep artifact bytes on the hub's own disk, a
|
||||
mounted NAS share, or your own cloud bucket. Switched at runtime from
|
||||
Settings → Storage, no redeploy, and previously used backends stay readable so
|
||||
a switch never orphans what you already had.
|
||||
- **Reference-only artifacts** — catalogue what you already hold without copying
|
||||
it. A 57 GB repository registers in seconds having downloaded nothing, and
|
||||
costs nothing against quota.
|
||||
|
||||
### Workloads
|
||||
|
||||
- GPU-aware scheduling with eviction and sweeps, managed volumes with a file
|
||||
browser and write API, notebooks, serving, and promote-a-result into the
|
||||
catalogue.
|
||||
|
||||
### Owl / AI operations
|
||||
|
||||
- Owl acts through **risk-classified tools** with indirect prompt-injection
|
||||
defences, and has deployed and configured Prometheus end to end unattended.
|
||||
- Owl affordances hide themselves when no model key is configured, rather than
|
||||
offering actions that cannot work.
|
||||
|
||||
### Security
|
||||
|
||||
- **Auth middleware rewritten.** The old public-route list matched by path
|
||||
prefix and ignored the HTTP method, which silently exempted every sub-path and
|
||||
every method beneath a public entry. Public routes are now exact
|
||||
(method, prefix) pairs. Regression tests lock this in and run on every push.
|
||||
- Per-node client certificates issued by the hub's own CA, with the agent
|
||||
pinning the CA fingerprint; encrypted secret store for tokens and credentials.
|
||||
|
||||
### Project
|
||||
|
||||
- CI on every push: build, vet and the full Go test suite, plus a frontend
|
||||
type-check and a check that fails the build if a React component nothing
|
||||
imports would ship unreachable.
|
||||
- Public documentation site at **[aegis.jerxie.com](https://aegis.jerxie.com)**,
|
||||
with an assistant that answers from this documentation — including the parts
|
||||
about what Aegis cannot do yet.
|
||||
|
||||
### Known gaps
|
||||
|
||||
Stated plainly rather than left to be discovered: single operator account (no
|
||||
multi-user or RBAC), no control-plane HA, training not yet modelled as a managed
|
||||
workload, and scale validated at nine nodes.
|
||||
|
||||
## [1.1.0] — 2026-05-23
|
||||
|
||||
### Gateway — Protection
|
||||
|
||||
@@ -5,7 +5,7 @@
|
||||
<h1 align="center">Aegis</h1>
|
||||
|
||||
<p align="center">
|
||||
<strong>Self-hosted Envoy gateway · AI threat analysis · TLS automation · Real-time dashboard</strong>
|
||||
<strong>Run AI on hardware you control.</strong>
|
||||
</p>
|
||||
|
||||
<p align="center">
|
||||
@@ -15,7 +15,13 @@
|
||||
</p>
|
||||
|
||||
<p align="center">
|
||||
Aegis sits between the internet and your services — one container that controls Envoy Proxy in real time, blocks threats automatically, manages TLS certificates, and lets you chat with your gateway through an AI assistant.
|
||||
One control plane from the bare metal up — provisioning, networking, Kubernetes, models and serving.<br />
|
||||
Fully air-gapped when you need it. Your model weights stay on your storage.<br />
|
||||
With an AI operator that does the routine work instead of describing it.
|
||||
</p>
|
||||
|
||||
<p align="center">
|
||||
<a href="https://aegis.jerxie.com"><strong>aegis.jerxie.com</strong></a> — ask the assistant anything about the platform
|
||||
</p>
|
||||
|
||||
---
|
||||
@@ -24,6 +30,24 @@
|
||||
|
||||
---
|
||||
|
||||
## What Aegis is
|
||||
|
||||
Every organisation that cannot put its data in someone else's cloud currently
|
||||
assembles an on-prem AI stack out of six tools — a mesh VPN, an ingress gateway,
|
||||
a Kubernetes distribution, a bare-metal provisioner, an artifact registry, a
|
||||
monitoring stack — plus a team to keep them in step. That team is the product
|
||||
they are actually buying, and it does not scale.
|
||||
|
||||
Aegis replaces the assembly. One hub container runs the control plane; a single
|
||||
agent enrols each machine. From there you provision bare metal, build clusters,
|
||||
catalogue models, and serve them — without any of it leaving your network.
|
||||
|
||||
> **Note on scope.** Earlier versions of this README described Aegis as an Envoy
|
||||
> gateway with AI threat analysis. That is now one subsystem of several. The
|
||||
> gateway still does everything it did; the product around it grew.
|
||||
|
||||
---
|
||||
|
||||
## Quick start
|
||||
|
||||
```bash
|
||||
@@ -36,151 +60,110 @@ curl -o envoy/envoy.yaml https://raw.githubusercontent.com/axieyangb/aegis/main/
|
||||
docker compose up -d
|
||||
```
|
||||
|
||||
Open **`http://localhost:8765`** — default login: `admin` / `changeme`.
|
||||
Open **`http://localhost:8765`** — default login `admin` / `changeme`.
|
||||
|
||||
> Edit `docker-compose.yml` and set `ADMIN_PASSWORD` before exposing to the network.
|
||||
> Set `ADMIN_PASSWORD` in `docker-compose.yml` before exposing this to anything.
|
||||
|
||||
On first boot Aegis seeds a working gateway baseline — HTTP (port 10080) and HTTPS (port 10443) listeners ready to accept filter chains.
|
||||
Then enrol a machine — one line on the target, no SSH keys to distribute and no
|
||||
inbound ports to open:
|
||||
|
||||
```bash
|
||||
curl -fsSL "http://<your-hub>:8765/install.sh?token=<join-token>" | sudo bash
|
||||
```
|
||||
|
||||
The node dials out to the hub, receives an identity signed by your hub's own CA,
|
||||
and appears in the fleet. Bare metal with a BMC can instead be powered on and
|
||||
installed from the UI.
|
||||
|
||||
**Requirements:** the hub needs any Linux host with Docker (2 vCPU / 2 GB is
|
||||
plenty). Nodes need Linux with systemd, x86_64 or arm64. GPUs are detected and
|
||||
scheduled on automatically. Internet access is optional after the first pull.
|
||||
|
||||
---
|
||||
|
||||
## Features
|
||||
## What it does
|
||||
|
||||
<table>
|
||||
<tr>
|
||||
<td align="center" width="50%">
|
||||
| | |
|
||||
|---|---|
|
||||
| **Fleet** | One-line enrolment, per-node mTLS identity, digest-verified agent upgrades with systemd auto-rollback, activity log |
|
||||
| **Networking** | WireGuard overlay mesh, STUN NAT traversal, automatic LAN-direct routing with mesh fallback, embedded fleet DNS for hosts *and* containers |
|
||||
| **Gateway** | Envoy with a full xDS control plane, SDS certificate management, automated TLS issuance, staged pending→apply config, rate limiting, 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 clusters via a reconciler — k3s and kubeadm, with air-gapped installs validated live |
|
||||
| **Models & data** | Catalogue models and datasets with versions, digests and lineage; pull from HuggingFace once and stage to nodes over the LAN; quotas per owner |
|
||||
| **Your storage** | Keep artifact bytes on your own NAS mount or cloud bucket, switched at runtime from the UI. Reference-only artifacts catalogue what you already have without copying it |
|
||||
| **Workloads** | GPU-aware scheduling, managed volumes with a file browser, notebooks, serving, promote-a-result-into-the-catalogue |
|
||||
| **AI operations** | Owl, an assistant that acts through risk-classified tools with prompt-injection defences — it has deployed and configured Prometheus end to end, unattended |
|
||||
| **Protection** | AI threat analysis over gateway access logs, automatic blocking, patrol sweeps, notifications |
|
||||
|
||||
**🦉 Owl AI Assistant**
|
||||
---
|
||||
|
||||
|
||||
## What makes it different
|
||||
|
||||
Chat with your gateway in plain English. Owl analyses traffic, surfaces threats, and can configure your gateway end-to-end — clusters, certs, filter chains — from a single prompt.
|
||||
Most of the above has a good commercial equivalent. These do not:
|
||||
|
||||
</td>
|
||||
<td align="center" width="50%">
|
||||
- **Air-gap is the default, not an enterprise tier.** A 5-node HA Kubernetes
|
||||
cluster installed with no internet access at all, validated live. It falls out
|
||||
of the architecture: the hub reaches the internet once, every node pulls from
|
||||
the hub.
|
||||
- **An AI operator that operates.** Owl changes infrastructure through scoped,
|
||||
risk-classified tools — not a chat box that summarises a dashboard.
|
||||
- **Bare metal to running model in one control plane.** The seams between MAAS,
|
||||
Rancher and Harbor are where an ops team currently lives.
|
||||
- **Your weights never touch our disk.** We catalogue; your storage holds. A
|
||||
57 GB repository can be registered in seconds having downloaded nothing.
|
||||
- **One transfer, then LAN.** A node with no internet access still gets a
|
||||
57.9 GB model.
|
||||
|
||||
**🛡 Envoy Gateway Control**
|
||||
---
|
||||
|
||||
|
||||
## What it does not do yet
|
||||
|
||||
Visual editor for listeners, filter chains, and clusters. Changes are validated and pushed live to Envoy via xDS — no restarts, no YAML files.
|
||||
Published deliberately, because finding this out later is worse:
|
||||
|
||||
</td>
|
||||
</tr>
|
||||
<tr>
|
||||
<td align="center" width="50%">
|
||||
|
||||
**🔍 IP Intelligence**
|
||||
|
||||
|
||||
|
||||
Every IP auto-profiled: geolocation, ASN, VPN/Tor detection, AbuseIPDB score, and full request history. AI patrol sweeps run in the background and auto-block threats.
|
||||
|
||||
</td>
|
||||
<td align="center" width="50%">
|
||||
|
||||
**🔒 TLS Automation**
|
||||
|
||||
|
||||
|
||||
ACME auto-renewal (Let's Encrypt, ZeroSSL), HTTP-01 & DNS-01 challenges, and a built-in Local CA for internal services — all pushed directly to Envoy SDS.
|
||||
|
||||
</td>
|
||||
</tr>
|
||||
<tr>
|
||||
<td align="center" width="50%">
|
||||
|
||||
**🔔 AI Patrol & Alerts**
|
||||
|
||||
|
||||
|
||||
Scheduled AI sweeps classify traffic around the clock. Blocks and anomalies are pushed to Telegram, Discord, Slack, or webhook.
|
||||
|
||||
</td>
|
||||
<td align="center" width="50%">
|
||||
|
||||
**📱 Mobile-ready**
|
||||
|
||||
|
||||
|
||||
Full dashboard and Owl chat from any device. Ask Owl what happened in the last two hours — it triages threats, blocks IPs, and confirms — all from your phone.
|
||||
|
||||
</td>
|
||||
</tr>
|
||||
<tr>
|
||||
<td align="center" width="50%">
|
||||
|
||||
**🚧 Protection & Maintenance**
|
||||
|
||||
Styled 403/429 block pages with a customisable contact message — replace bare status codes with a branded denial page. Maintenance mode puts any listener (or a single SNI domain) into a 503 page while leaving all other services online. Preview pages inline before pushing to Envoy.
|
||||
|
||||
</td>
|
||||
<td align="center" width="50%">
|
||||
|
||||
**⚡ Per-Domain Maintenance via Owl**
|
||||
|
||||
Ask Owl to put a single domain into maintenance while the rest of your services stay live: Owl calls `gateway_list_listeners` to discover the SNI, then targets only that filter chain — no manual JSON editing required.
|
||||
|
||||
</td>
|
||||
</tr>
|
||||
</table>
|
||||
| Gap | Where it stands |
|
||||
|---|---|
|
||||
| **Multi-user / RBAC** | Single operator account. Ownership already threads through the data model; the identity layer on top is the next major workstream. |
|
||||
| **Control-plane HA** | One hub, one database. Workloads survive a hub outage; management does not. |
|
||||
| **Managed training runs** | The GPU scheduler, artifact capture and volumes all exist, but training is not yet modelled as a workload — a run is still started by hand. |
|
||||
| **Scale** | Validated at nine nodes. We quote that number rather than implying a larger one. |
|
||||
|
||||
---
|
||||
|
||||
## Architecture
|
||||
|
||||
```
|
||||
Internet ──▶ Envoy Proxy ──▶ Your services
|
||||
│
|
||||
gRPC xDS (port 18000)
|
||||
│
|
||||
┌────▼─────┐
|
||||
│ Aegis │ port 8765
|
||||
│ │
|
||||
│ xDS CP │ controls Envoy live
|
||||
│ Analytics│ reads Envoy ALS logs
|
||||
│ AI Engine│ classifies IPs
|
||||
│ Cert Mgr │ ACME + Local CA → Envoy SDS
|
||||
│ Dashboard│ web UI + REST API
|
||||
└──────────┘
|
||||
```
|
||||
Aegis runs as two containers — the hub (control plane, API, UI) and Envoy
|
||||
(data plane, driven entirely over xDS). Agents on each node dial **out** to the
|
||||
hub over a persistent connection, so no node needs an inbound port and nothing
|
||||
needs an SSH key distributed to it.
|
||||
|
||||
`linux/amd64` and `linux/arm64` — runs on x86 servers, Raspberry Pi, Synology NAS, and Apple Silicon.
|
||||
See [deployment architectures](docs/deployment-architectures.md) for direct
|
||||
exposure, VPS relay and Cloudflare Tunnel topologies, and
|
||||
[sequence diagrams](docs/sequence-diagrams.md) for the request paths.
|
||||
|
||||
---
|
||||
|
||||
## Configuration
|
||||
|
||||
| Variable | Default | Description |
|
||||
|---|---|---|
|
||||
| `PORT` | `8765` | Dashboard + API port |
|
||||
| `XDS_PORT` | `18000` | Envoy gRPC xDS port |
|
||||
| `DATA_DIR` | `/data` | Persistent data directory |
|
||||
| `ADMIN_USERNAME` | `admin` | Admin username |
|
||||
| `ADMIN_PASSWORD` | `aegis` | Admin password — **change this** |
|
||||
| `AUTH_ENABLED` | `true` | Require login |
|
||||
| `BLOCK_ENABLED` | `true` | Enable auto IP blocking |
|
||||
| `NODE_ID` | `home` | Envoy node ID (must match envoy.yaml) |
|
||||
|
||||
Data is persisted at `/data/aegis.db` (SQLite). Mount a volume to keep data across container updates.
|
||||
Everything is environment variables in `docker-compose.yml`. See
|
||||
[configs/](configs/) for annotated examples and
|
||||
[envoy-config.md](docs/envoy-config.md) for the gateway bootstrap.
|
||||
|
||||
---
|
||||
|
||||
## Docs & Tutorials
|
||||
## Docs & tutorials
|
||||
|
||||
- [Getting started](docs/getting-started.md)
|
||||
- [Envoy configuration](docs/envoy-config.md)
|
||||
- [AI setup — Owl chat + threat analysis](docs/ai-setup.md)
|
||||
- [Notifications — Telegram, Discord, webhooks](docs/notifications.md)
|
||||
- [AI setup](docs/ai-setup.md) — connecting Owl to a model provider
|
||||
- [Notifications](docs/notifications.md)
|
||||
- [Deployment architectures](docs/deployment-architectures.md)
|
||||
|
||||
### Tutorial series: Exposing a service with Aegis
|
||||
### Tutorial series: exposing a service with Aegis
|
||||
|
||||
| # | Tutorial | Description |
|
||||
|---|---|---|
|
||||
| 01 | [Local HTTPS with a whoami service](docs/tutorials/01-whoami-local-https.md) | Configure the gateway manually through the UI |
|
||||
| 02 | [Configure the Gateway with Owl AI](docs/tutorials/02-whoami-ai-setup.md) | Same setup — let Owl AI do the configuration from a single prompt |
|
||||
| 03 | [Understanding the Dashboard](docs/tutorials/03-understanding-the-dashboard.md) | Read live traffic data and analyse request patterns with Owl |
|
||||
| 04 | [AI-Driven Protection](docs/tutorials/04-ai-driven-protection.md) | Use Owl to disable a service under attack and bring it back |
|
||||
1. [Local HTTPS with whoami](docs/tutorials/01-whoami-local-https.md)
|
||||
2. [AI setup](docs/tutorials/02-whoami-ai-setup.md)
|
||||
3. [Understanding the dashboard](docs/tutorials/03-understanding-the-dashboard.md)
|
||||
4. [AI-driven protection](docs/tutorials/04-ai-driven-protection.md)
|
||||
|
||||
### Videos
|
||||
|
||||
@@ -211,13 +194,14 @@ One prompt to Owl AI sets up the cluster, issues a certificate, and wires the fi
|
||||
|
||||
Distributed as a compiled binary. Source code is proprietary. See [LICENSE](LICENSE).
|
||||
|
||||
Community tier is **free forever**. Pro unlocks unlimited notification channels, longer log retention, and unlimited AI patrol sweeps.
|
||||
Community tier is **free forever**. Pro unlocks unlimited notification channels,
|
||||
longer log retention, and unlimited AI patrol sweeps.
|
||||
|
||||
---
|
||||
|
||||
## About
|
||||
|
||||
Built by **Jerry Xie** — formerly network security at Palo Alto Networks, now Senior Software Engineer specialising in identity, distributed cloud, Kubernetes, and AI. Aegis started as a home lab project and grew into a product.
|
||||
|
||||
**Issues & feature requests:** [GitHub Issues](https://github.com/axieyangb/aegis/issues)
|
||||
**Enterprise / custom integrations:** [yyangxie@gmail.com](mailto:yyangxie@gmail.com)
|
||||
Built and run on a real fleet — nine nodes spanning LAN, cloud and overlay-only,
|
||||
managed through this platform daily. Questions are welcome at
|
||||
[aegis.jerxie.com](https://aegis.jerxie.com), where an assistant answers from
|
||||
this documentation, including the parts about what Aegis cannot do yet.
|
||||
|
||||
+110
@@ -0,0 +1,110 @@
|
||||
# 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](https://aegis.jerxie.com), where an assistant answers from
|
||||
this documentation — including the parts about what is missing.
|
||||
+16
-1
@@ -4,7 +4,17 @@ services:
|
||||
container_name: aegis
|
||||
restart: unless-stopped
|
||||
ports:
|
||||
- "8765:8765"
|
||||
- "8765:8765" # UI + API
|
||||
# Agents dial IN to this port over mTLS. Without it published, a node can
|
||||
# be issued a join token and will never be able to use it — the one-line
|
||||
# installer completes and the node never appears. Only needed if you are
|
||||
# enrolling machines; a single-host gateway install can drop it.
|
||||
- "8766:8766" # agent control channel (mTLS)
|
||||
# Overlay mesh. Needed for nodes that are not on this LAN; harmless if
|
||||
# unused. UDP, and both are required — STUN discovers the path, WireGuard
|
||||
# carries the traffic.
|
||||
- "3478:3478/udp" # STUN — NAT traversal
|
||||
- "51820:51820/udp" # WireGuard overlay
|
||||
volumes:
|
||||
- aegis_data:/data
|
||||
environment:
|
||||
@@ -13,6 +23,11 @@ services:
|
||||
- AUTH_ENABLED=true
|
||||
- BLOCK_ENABLED=true
|
||||
- NODE_ID=home
|
||||
# The address agents should dial back on. Leave blank for a gateway-only
|
||||
# install; set it to a hostname or IP the nodes can actually reach before
|
||||
# enrolling anything, or the token will carry an address that does not
|
||||
# resolve from the node.
|
||||
- AGENT_HOST=
|
||||
|
||||
envoy:
|
||||
image: envoyproxy/envoy:v1.35-latest
|
||||
|
||||
Vendored
BIN
Binary file not shown.
@@ -1,6 +1,6 @@
|
||||
# Deployment Architectures (Exposing Envoy)
|
||||
|
||||
To make Aegis and Envoy accept real public traffic, you need to expose Envoy's ports (Host `80`/`443`) to the internet. Depending on your network setup (home lab, static IP, CGNAT, VPS), you should choose one of the two main architectures below.
|
||||
To make Aegis and Envoy accept real public traffic, you need to expose Envoy's ports (Host `80`/`443`) to the internet. Depending on your network setup (on-prem LAN, static IP, CGNAT, VPS), you should choose one of the two main architectures below.
|
||||
|
||||
---
|
||||
|
||||
|
||||
Reference in New Issue
Block a user