Don’t take our word.Run them yourself.
The same four agents from our studio — behind a real API. Run one live in the browser, or curl the endpoint. Real reasoning loops on sample data, recorded and replayable, never a mockup.
Plain-English question → real SQL → self-corrects → answer with the actual figures.
Or call it yourself
curl -X POST https://deviaisolutions.vercel.app/api/agents/sql \
-H 'content-type: application/json' \
-d '{"input": "Which billing country has the most invoices, and how many?"}'Same endpoint the console uses. Rate-limited, because every run executes on one real Claude subscription — no API key, no per-call billing. Want a dedicated, unthrottled API on your own data? That’s the product.
Pick an agent, tweak the input, and hit run. You’ll see the real reasoning loop — plan, call a tool, read the result, self-correct — not a canned reply.
01 — Keyless reasoning
Claude, no API key
The agents reason through the Claude CLI on a real subscription — no API key, no per-call billing. That’s also why they can’t run on serverless: they need the CLI.
02 — Self-hosted worker
A small guarded server
A zero-dependency Node worker runs the agents: one at a time, rate-limited, timed out, and key-protected — so a public console can’t drain a single subscription.
03 — Honest fallback
Always-on proof
If the worker is offline or a run runs long, the console replays a real recorded run and says so plainly. You always see genuine agent output — live when we can, recorded when we can’t.
/api/agents/sqlEnglish question → SQL → answer with real figures/api/agents/ragGrounded Q&A over documents — cites or refuses/api/agents/researchLive web research → sourced brief/api/agents/recruitScreen a pool, enforce deal-breakers in code/api/agents/healthIs the live worker up? (drives the fallback)Every endpoint returns the agent’s final answer plus the full trace of its reasoning and tool calls — the same recording the console replays. Rate-limited on the shared demo; unthrottled and on your own data when we build it for you.