Quickstart
This guide gets a local Aionis Runtime and product API running on your machine, then shows how to connect SDK, HTTP, MCP, AIFS, or native plugins. The goal is to complete one governed Agent loop:
observe -> guide -> agent action -> feedback -> measure -> snapshotAionis is local-first today. It is designed to sit beside your Agent host and compile governed context for that host.
Prerequisites
| Requirement | Why it matters |
|---|---|
| Node.js 22+ | The Lite Runtime uses the Node SQLite runtime. |
| An embedding key | Not needed for the first-value demo. Recall-backed guide flows need embeddings; OpenAI-compatible embeddings are the default path. |
| A local shell | The quickstarts start the Runtime and run SDK scripts locally. |
The Lite Runtime defaults to loopback and local unauthenticated use, which keeps the first install fast for local Agent development.
One-command Install
The fastest path is the installer package:
npx @aionis/create@latestThe installer clones the Runtime, installs Runtime dependencies, writes a local
.env, runs the Runtime build check, and starts the no-key first-value demo.
SDK, MCP, and Claude Code integrations are published from their own package
repositories and connect to this Runtime through the product API.
For the full SDK quickstart with recall-backed guide output:
OPENAI_API_KEY="your-key" npx @aionis/create@latest --provider openai --quickstart sdkIf you already cloned the repository, use the local workflow below.
Optional Claude Code Plugin
If your first target is Claude Code, install Runtime into a side directory:
npx @aionis/create@latest .aionis-runtimeThen start Runtime:
cd .aionis-runtime
npm run -s lite:startNow run Claude Code from your project:
claudeInside Claude Code, install the plugin:
/plugin marketplace add https://github.com/ostinatocc/aionis-claude-code
/plugin install aionis@aionis-claude-code
/aionis:onboardClaude Code will receive Aionis context before prompts and Aionis will record execution evidence after tool use. The plugin path installs user-level hooks and MCP tools, so future projects do not need manual hook setup.
Local Repository Setup
git clone https://github.com/ostinatocc/Aionis.git
cd Aionis
npm installRun the first-value demo without an embedding key:
npm run -s runtime:demo:first-valueThen configure embeddings:
export EMBEDDING_PROVIDER="openai"
export OPENAI_API_KEY="your-openai-key"MiniMax is also supported:
export EMBEDDING_PROVIDER="minimax"
export MINIMAX_API_KEY="your-minimax-key"Run the SDK quickstart:
npm run -s runtime:quickstart:sdkThat quickstart exercises the product path:
rememberwrites ordinary scoped memory.observeSteprecords execution evidence.guideForRolecompiles governed Agent context.feedbackFromOutcomeattributes the result to memory that was actually exposed by the guide.measureRunreports whether memory helped or created risk.snapshotRuncreates an operator-facing replay view.
What Success Looks Like
A successful run prints a JSON result with fields like:
{
"prompt_preview": "AIONIS EXECUTION CONTEXT...",
"execution_context_contract": "aionis_execution_agent_context_v1",
"memory_use_receipt_visible": true,
"used_memory_ids": ["mem_..."],
"snapshot_visible": true
}The important part is the contract shape:
| Field | Meaning |
|---|---|
prompt_preview | The Agent-facing context compiled by Aionis. |
memory_use_receipt_visible | The guide includes a compact audit of what was used or suppressed. |
used_memory_ids | The host can attribute outcome only to memory exposed by that guide. |
snapshot_visible | Operator replay can explain what Aionis admitted and why. |
Choose Your First Loop
| If you are building | Start with | What it proves |
|---|---|---|
| First local proof | npm run -s runtime:demo:first-value | Outdated or low-authority retrieved memory is governed before prompt use |
| Claude Code | /plugin install aionis@aionis-claude-code | Plugin hooks inject context and record execution evidence automatically |
| Cursor or MCP-only hosts | npx @aionis/mcp@latest --base-url http://127.0.0.1:3001 --scope-from workspace --workspace-id-store user | MCP tools can record steps and compile context |
| TypeScript or Node Agent | npm run -s runtime:quickstart:sdk | remember -> guide -> feedback -> measure -> snapshot |
| Raw HTTP service | npm run -s runtime:quickstart:http | Public HTTP surface without SDK helpers |
| Multi-agent system | npm run -s runtime:quickstart:multi-agent | Shared execution memory across planner, worker, verifier, reviewer |
| External memory backend | npm run -s runtime:quickstart:memory-firewall | Govern Mem0/Zep/vector/markdown candidates |
| Operator debugging | npm run -s runtime:quickstart:flight-recorder | Read-only incident replay without prompt payload |
Source guide: AIONIS_QUICKSTART_MATRIX.md .
Minimal SDK Loop
import {
agentPromptFromGuide,
createAionisClient,
feedbackFromGuide,
} from "@aionis/sdk";
const aionis = createAionisClient({
baseUrl: "http://127.0.0.1:3001",
scope: "my-project",
});
await aionis.remember({
kind: "preference",
text: "Do not reuse failed migration branches as current instructions.",
});
const guide = await aionis.guide({
query_text: "Continue the migration from the current accepted path.",
context_mode: "compact_agent",
});
const prompt = agentPromptFromGuide(guide);
const result = await agent.run(prompt);
const usedMemoryIds = result.usedMemoryIds ?? [];
if (usedMemoryIds.length > 0) {
await aionis.feedback(feedbackFromGuide({
guide,
run_id: result.runId,
outcome: result.ok ? "positive" : "negative",
reason: result.summary,
used_memory_ids: usedMemoryIds,
used_surface: "use_now",
}));
}Common First-run Issues
| Symptom | Fix |
|---|---|
| Runtime cannot recall anything | Check EMBEDDING_PROVIDER and provider API key. |
| Guide has no useful history | Make sure tenant_id, scope, and Agent identity match between observe and guide. |
| Feedback is rejected | Aionis only accepts attribution to memory IDs exposed by that guide. |
| MCP client sees no tools | Confirm the MCP server points at the same Runtime base-url. |
| You want remote access controls | Run the managed-server configuration path with API-key or JWT auth settings. |
Next
- Follow the full Get Started path.
- Build with Execution Memory.
- Connect a TypeScript host with the SDK guide.
- Add Memory Firewall in front of Mem0 or a vector DB with the Memory Firewall guide.
- Use Claude Code or Cursor through MCP.