Auggy runs model turns and typed tools. Augments add memory, context, transports, skills, policy, and integrations without expanding the kernel or taking over the rest of your stack.
Scaffold a project, configure the provider you selected, validate it, and open creator chat from the same process.
Create the concierge project with the documented CLI path.
For a hosted provider, add the credential named in .env.example to .env. For Ollama, make sure the local service is running.
Doctor checks the project and provider before the runtime starts.
Small TypeScript agent runtimenpm i -g auggy
auggy create concierge
cd conciergeauggy doctor
auggy runAgent "concierge" is live.
Chat: http://localhost:8080/console/chat
Console: http://localhost:8080/console
Health: http://localhost:8080/health
Home: http://localhost:8080/Start from the compact chat-ready project. Install identity, memory, communication, or external tools independently; each augment contributes one coherent runtime capability.
Recognize returning visitors
Verify visitors with an emailed magic link, then issue a stable recognized-visitor token for later sessions. Pair that durable identity with memory to personalize future agent experiences.
auggy augment add visitorAuthSend, receive, and recover
Give Auggy an agent-native inbox for policy-gated sending, receiving, replies, and forwarding, with durable catch-up, creator review, and trust, recipient, rate, content, and audit controls.
auggy augment add agentMailRemember people and agents
Give Auggy peer-scoped memory for other agents, users, and recognized visitors, so relevant facts and context can return in later conversations.
auggy augment add layeredMemoryConnect existing tool servers
Import local or remote MCP tools with namespacing, allowlists, trust policy, lifecycle checks, and cloud diagnostics.
auggy augment add mcpA custom augment packages everything needed for one job—routes your app can call, tools your agent can use, access rules, and ordinary functions that do the work.
Build the workflow once, secure it in one place, and use it from both your app and your agent.
augments/sales/
├── index.ts mounts the routes and tools
├── routes.ts direct app and webhook operations
├── tools.ts model-callable agent actions
├── functions.ts ordinary TypeScript that does the work
├── auth.ts access rules used by routes and tools
└── schemas.ts validated inputs for each operationimport { defineAugment } from "auggy";
import { salesRoutes } from "./routes";
import { salesTools } from "./tools";
export default function sales() {
return defineAugment({
name: "sales",
type: "custom",
httpRoutes: salesRoutes,
tools: salesTools,
});
}Use the creator console to test conversations, expand visible tool calls, inspect live integrations and capabilities, and tune the supported public discovery setting.
The 0.5 console is deliberately not a full configuration editor, raw model trace viewer, process manager, or multi-agent fleet dashboard.
Explore the console
Read rendered answers, assistant errors, and a copyable Markdown transcript from the running agent.
Expand validated arguments, results, category, and completion state beside the reply that used them.
Exercise creator, anonymous visitor, and verified visitor posture without pretending identity is model judgment.
Inspect mounted capabilities, endpoints, auth warnings, and toggle public developer discovery when you need it.
Run Auggy locally while you build, then send the same project through its guided Railway deploy path when it needs a public URL and durable state.
# Run in the foreground on this machine
auggy doctor
auggy run
# Or deploy the same agent project to Railway
railway login
auggy deployconcierge is live on Railway.
Chat: https://concierge-production-xxxx.up.railway.app/console/chat
Console: https://concierge-production-xxxx.up.railway.app/console
Health: https://concierge-production-xxxx.up.railway.app/health
Home: https://concierge-production-xxxx.up.railway.app/
Manage it:
auggy logs
auggy deploy --yes Redeploy
Details:
Service: concierge
Volume: concierge-data (mounted at /app/data)Preflight
Config, environment, dependencies, ports, and cloud-safe augment posture.
Persistent state
One service and one volume mounted at /app/data for durable runtime data.
Public verification
A generated domain, bounded deployment wait, and public /health check.
Auggy's kernel is deliberately small. We integrated findings from ALARA, AIOS, Letta, OpenAI, and other agent-runtime research into a lean architecture that reduces wrong tool selection, preserves the most important context by evicting lower-priority material first, and enforces permissions in code before tools run.
Impact on Auggy
Auggy exposes only tools from mounted augments and structurally removes disallowed tools. The model has fewer wrong choices, tool schemas consume less context, and hidden capabilities cannot be prompt-injected back into view.
Research basis
Impact on Auggy
Auggy allocates a bounded prompt, orders typed context by priority, and drops lower-priority blocks when space runs out. More of the window remains available for the conversation and the evidence needed now.
Impact on Auggy
Trust, schemas, rate limits, approvals, and per-augment tool caps run in code around each action. Safety does not depend on the model remembering a sentence, and runaway work is bounded where it starts.
Impact on Auggy
Auggy has one extension primitive: the augment. Fewer overlapping abstractions mean less framework surface to learn, fewer hidden hooks to debug and secure, and capabilities that can evolve without rewriting the turn loop.
Start with a successful local turn so every later change has a working baseline. Add a built-in, write one custom capability, inspect it, then deploy it.
