You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: README.md
+8-5Lines changed: 8 additions & 5 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -23,13 +23,13 @@ AEGIS is a framework for building **personal AI agents** that remember everythin
23
23
-**Standalone** — Clone, configure, deploy. Full agent in minutes.
24
24
-**As a dependency** — `npm install @stackbilt/aegis-core` and extend with your own routes, scheduled tasks, executors, and MCP tools via `createAegisApp()`.
25
25
26
-
The production instance runs 26 scheduled tasks, has executed 236+ autonomous coding sessions, and costs $0/month to host (Cloudflare Workers free tier + Workers AI for inference).
26
+
The reference deployment runs the same scheduled-task framework and is designed for $0/month hosting on the Cloudflare Workers free tier, with Workers AI as the base inference path.
27
27
28
28
Built on Cloudflare Workers for edge-native deployment. Zero cold starts. Global distribution. Pay-per-request economics.
29
29
30
30
### Core Capabilities
31
31
32
-
-**Cognitive Kernel** — Multi-model dispatch (Claude, Groq, Workers AI) with procedural memory routing. The right model for the right task, automatically.
32
+
-**Cognitive Kernel** — Workers AI-first dispatch with optional Claude and Groq executors, plus procedural memory routing. The right model for the right task, automatically.
33
33
-**Multi-Tier Memory** — Episodic (what happened), semantic (what matters), procedural (what works), narrative (the story arc). Memory consolidates, decays, and strengthens over time.
34
34
-**Autonomous Goals** — Set goals with standing orders and let AEGIS pursue them on a schedule. Progress tracked, blockers surfaced, results reported.
35
35
-**Dreaming Cycle** — Nightly self-reflection over conversation history. Discovers patterns, proposes improvements, proposes new tools, consolidates knowledge. PRISM synthesis finds cross-domain connections.
@@ -71,7 +71,7 @@ Talk to the same deployment from a terminal:
As of 0.8.0, the public package ships as a self-sufficient Cloudflare Worker app, not just a kernel library. A fresh checkout can build and deploy the browser console, text chat, voice adapter, scheduled tasks, MCP server, and core API routes without private Stackbilt UI code.
32
+
33
+
The standalone Worker uses these bindings:
34
+
35
+
| Binding | Type | Purpose |
36
+
|---------|------|---------|
37
+
|`ASSETS`| Workers static assets | Serves the Vite-built web console from `web/public/`|
38
+
|`DB`| D1 database | Stores conversations, memory, goals, agenda, tools, and task state |
39
+
|`AI`| Workers AI | Base chat, voice, classification fallback, and low-cost summarization |
1.`npm run build:ui` builds `web/src/ui/` into `web/public/`.
46
+
2.`wrangler.toml.example` wires `ASSETS`, `DB`, `AI`, `CHAT_SESSION`, and `AegisVoiceAdapter`.
47
+
3.`npm run deploy` builds the UI and runs `wrangler deploy`.
48
+
4. Visiting the Worker root serves the embedded console through `ASSETS`.
49
+
5.`/api/*`, `/health`, `/chat/ws`, and `/agents/*` run through the Worker before static asset fallback.
50
+
51
+
All interactive routes use the same `AEGIS_TOKEN` bearer-token boundary. The browser voice route also accepts the token through the voice call query path so the Agents SDK adapter can authenticate the session before model routing.
52
+
53
+
Package consumers can still import `createAegisApp()` and compose their own routes, scheduled tasks, executors, and MCP tools. The standalone app is the reference assembly; the package exports the primitives needed to build variants.
54
+
25
55
## The dispatch cycle
26
56
27
57
Every user message follows the same path through the kernel:
@@ -53,10 +83,10 @@ AEGIS supports multiple AI backends, selected per-request based on intent and le
This script is the launch-safe path for recording or presenting AEGIS 0.8.0. It demonstrates the two public surfaces that now work from a standalone checkout: the embedded browser console and the package CLI.
4
+
5
+
## Goal
6
+
7
+
Show that a developer can install AEGIS, deploy a self-contained Cloudflare Worker, use the browser console, and talk to the same agent from a terminal without private Stackbilt infrastructure.
8
+
9
+
## Preflight
10
+
11
+
Run these from `web/` before recording:
12
+
13
+
```bash
14
+
npm install
15
+
npm run build:ui
16
+
npm run typecheck
17
+
cp wrangler.toml.example wrangler.toml
18
+
npx wrangler deploy --dry-run
19
+
```
20
+
21
+
The dry run should show the `ASSETS`, `DB`, `AI`, `CHAT_SESSION`, and `AegisVoiceAdapter` bindings.
22
+
23
+
For the published package smoke:
24
+
25
+
```bash
26
+
npm pack @stackbilt/aegis-core@0.8.0
27
+
node package/cli/aegis.mjs --help
28
+
```
29
+
30
+
## Recording Flow
31
+
32
+
1. Open the Worker root and authenticate with `AEGIS_TOKEN`.
33
+
2. Show the health summary in the embedded console.
34
+
3. Send one grounded prompt:
35
+
36
+
```text
37
+
What runtime and bindings are you using for this session?
38
+
```
39
+
40
+
4. Send one continuity prompt:
41
+
42
+
```text
43
+
Summarize what changed in AEGIS 0.8.0 using only release metadata or health data you can access.
44
+
```
45
+
46
+
5. Open the voice panel and establish a call through `/agents/aegis-voice-adapter/operator`.
7. Ask the same release-summary prompt from the terminal to show that the browser and CLI are talking to the same deployed agent.
54
+
8. End by showing [proof-of-work-0.8.0.md](proof-of-work-0.8.0.md).
55
+
56
+
## Trust Guardrails
57
+
58
+
- Keep the demo to health, release, configuration, memory, agenda, and directly available deployment facts.
59
+
- Do not ask for live GitHub, npm, analytics, or production metrics unless the integration is configured and the response includes the tool/source path.
60
+
- Treat Claude and Groq as optional upgrades. The base demo should work with `AEGIS_TOKEN`, D1, Workers AI, Durable Objects, and static assets.
61
+
- If the agent cannot verify a fact from its configured surfaces, the correct answer is a clarification or limitation, not a guess.
0 commit comments