AskElira is an autonomous build platform that turns a plain-text goal into working code through a multi-agent pipeline. Each agent researches, builds, audits, and approves — producing verified output with zero manual intervention.
Every goal flows through six stages in sequence:
| Step | Agent | Role |
|---|---|---|
| 1 | Alba | Research — web search, pattern matching, risk analysis |
| 2 | Vex Gate 1 | Audit Alba's research for quality and completeness |
| 3 | David | Build — generates code, configs, and file artifacts |
| 4 | Vex Gate 2 | Audit David's output for correctness and security |
| 5 | Elira | Final review — approve, reject, or request iteration |
| 6 | Finalize | Persist results, notify, advance to next floor |
If Vex or Elira rejects, the pipeline loops back to Alba (up to 5 iterations).
# 1. Install
npm install
# 2. Initialize workspace (creates ~/askelira/ with SOUL.md, AGENTS.md, TOOLS.md)
npx askelira init
# 3. Onboard with OpenClaw gateway
openclaw onboard
# 4. Copy env and fill in your keys
cp .env.example .env
# 5. Start dev server
npm run devCopy .env.example to .env. Key variables:
| Variable | Description |
|---|---|
ANTHROPIC_API_KEY |
Claude API key for all agent reasoning |
SEARCH_PROVIDER |
Search backend: auto, tavily, brave, perplexity |
TAVILY_API_KEY |
Tavily search API key (used when SEARCH_PROVIDER=tavily or auto) |
BRAVE_SEARCH_API_KEY |
Brave Search API key (fallback when Tavily unavailable) |
OPENCLAW_GATEWAY_URL |
WebSocket URL for OpenClaw gateway (ws://127.0.0.1:18789) |
OPENCLAW_GATEWAY_TOKEN |
Auth token for gateway connection |
AGENT_ROUTING_MODE |
Agent routing strategy: gateway, direct, gateway-only |
TELEGRAM_BOT_TOKEN |
Telegram Bot API token for notifications |
TELEGRAM_CHAT_ID |
Telegram chat ID to receive notifications |
POSTGRES_URL |
PostgreSQL connection string |
NEXTAUTH_SECRET |
NextAuth session encryption key |
NEXTAUTH_URL |
Base URL for auth callbacks (http://localhost:3000) |
CRON_SECRET |
Secret for authenticating cron/internal API calls |
Set AGENT_ROUTING_MODE to control how agent calls are dispatched:
| Mode | Behavior |
|---|---|
gateway |
Prefer OpenClaw gateway, fall back to direct Anthropic API on failure (default) |
direct |
Always call Anthropic API directly, bypass gateway entirely |
gateway-only |
Require gateway — fail hard if gateway is unavailable |
The gateway client connects with Ed25519 device identity signing, circuit breaker (3 failures / 60s window opens for 300s), and auto-reconnect with exponential backoff.
Set SEARCH_PROVIDER to control Alba's web research:
| Provider | Notes |
|---|---|
auto |
Try Tavily first, fall back to Brave, then Perplexity |
tavily |
Tavily Search API only |
brave |
Brave Search API only |
perplexity |
Perplexity API only |
When TELEGRAM_BOT_TOKEN and TELEGRAM_CHAT_ID are set, AskElira sends notifications on:
- Gateway connected — successful WebSocket handshake
- Floor live — a floor starts its agent pipeline
- Goal met — all floors approved and finalized
- Floor blocked — a floor exceeded max iterations
- Circuit breaker open — gateway degraded after repeated failures
- Node.js 22.16+
- OpenClaw gateway running locally (
openclaw onboardto set up) - At least one search API key (Tavily or Brave) for Alba research
- PostgreSQL for goal/floor persistence
AGPL-3.0-or-later