Frequently asked questions.

Vulcan is a production marketplace for 17 AI Operators across five layers — Intelligence, Data, Communication, Experience, and Orchestration. Every agent owns a business outcome, every demo is live, and every refusal is typed. Try it free from the catalog or jump straight into the Agent Workbench.

Most marketplaces sell implementation names ("Web Scraper", "Email Agent"). Vulcan sells outcomes — a CFO doesn't buy a "NPV calculator", she buys a Capital Intelligence Agent that ships a board-ready memo. Every demo is real, runs against the same backend the production API will, and returns a typed envelope so you can integrate without surprises.

All 17 agents are live and shipping production responses today. The capability sprint timeline is published on the /updates page — every release for the last 10 days has touched at least one agent.

4 free runs per agent per day, per IP, resetting at UTC midnight. Purchased agents bypass the limit with up to 1,000 runs/day per sandbox key. The runs counter is shown live on every demo and updates the moment you fire a call.

On checkout success, three things happen in real time:

1. Payment confirmation (Razorpay test mode today).
2. A sandbox API key is issued from /v1/keys/issue and displayed with the endpoint + curl snippet.
3. A "Run in browser" button fires a live POST against the agent so you see the typed response inline — proof the key works before you leave the page.

Yes — the Workbench shows every agent, with a status badge for purchased ones. The Performance and Analytics views are filtered to your purchased agents only.

Every agent has a dedicated setup page at /workflows/<slug>/setup with five sections: the endpoint, your API key, copy-pasteable snippets for cURL / Python / JavaScript / MCP config, the sample payload, and the response envelope shape. Open one from the catalog.

Three steps. (1) Download vulcan-agent.mjs — a single Node 18+ file with zero dependencies. (2) Add the block from the agent's Setup · MCP tab to ~/Library/Application Support/Claude/claude_desktop_config.json (or Cursor → Settings → MCP servers) and point it at the absolute path. (3) Restart the client. A new tool run_<slug> appears wired to the endpoint with your bearer token.

All edge cases return the typed envelope — never a 500. Shape:

{
  "ok": false,
  "refusal": {
    "reason_code": "oversized_input",
    "human_message": "Inbox sample exceeds the 10-email cap.",
    "suggested_action": "Send 1–10 emails."
  },
  "request_id": "req_3f2c…"
}

Every agent owns a finite enum of reason_codes so your client can branch on them like any other typed error.

agents.aidesktop.in. Single FastAPI mono-service behind nginx, running pm2 on a hardened VPS. SSL via Let's Encrypt with auto-renew. Per-IP rate limiting (slowapi), structured request-ID logs, HMAC-signed bearer tokens.

The marketplace and agents-backend run as separate pm2 processes with independent restart policies. AnswerBot has its own backend at support.aidesktop.in. No agent talks to another by HTTP today — the only cross-agent path is the Executive Operations Agent's plan + execute mode (sandbox stubs).

Open the Contact page with the agent slug, the request_id from the response, and the input that triggered it. We treat 500s and untyped errors as P0 because they violate the envelope guarantee.