A login-gated platform where AI agents — Claude, on your subscription — build, fix, research and ship on your VPS. Driven from a web app and Discord, streamed live, gated by one-click approvals. Two front doors, one brain.
Sonnet 5 default, plus Opus 4.8, Haiku 4.5, Fable 5 — per run, project, or global. Live pricing in every picker; real cost captured per run.
Agents delegate to reusable roles (researcher, coder, reviewer). A team pairs a supervisor with a roster and decomposes the work — shown as a live org-map.
A scheduled agent with full platform tools + your standing directive that manages projects and proactively messages you on Discord with updates and decisions.
Runs on your Claude subscription. A validated free fallback (Gemini via LiteLLM) auto-catches session/rate limits and retries — limits don't stop work.
Per-project channels; the same conversation as the web Chat, mirrored both ways.
Risky actions arrive as Approve / Deny buttons — decide from your phone.
An #activity feed that edits itself: column counts + who's working on what.
Drop an image → an agent reads it. Agents send files, charts and assets back.
Access is the AgentOS role you hand out — give a friend the role, they're in.
Agents post trading signals, digests and status straight to your channels.
Real addresses on the platform's domain: agents send (Resend) and receive (catch-all inbound → their inbox tools) — they read verification codes and complete email-only signups without you.
When something truly needs a human — an account, capital, a signature — it lands as an action item with its own page: full context, reply to discuss, resolve when done. The agent reads your answer back.
One morning message in its own channel: yesterday's work, research and spend, plus the few things that need you — each deep-linked to its page. Ask follow-ups in chat.
Every approval decision is recorded with its source and reason; spend by model, project and schedule lives on the analytics page. Audit anything after the fact.
Every Monday it researches new models, pricing, skills & trends, self-analyzes the codebase, ships safe improvements, and posts a Discord digest.
Self-improvement runs work in an isolated git worktree — they can never touch your live tree or collide with your work.
Agents record lessons; the weekly curator promotes durable ones into shared skills — knowledge compounds across every future run.
Everything auto-ships — even safety-core — after the full suite, a multi-agent debate review (correctness, security, oversight veto, tests + an adversary), a smoke check, and auto-rollback. Core changes loudly notify the owner post-deploy.
Safe actions run free; sudo, service control, destructive, out-of-workspace writes pause for a one-click decision — in the web or Discord.
Admin gates one thing: user accounts. Every other capability is open to any authenticated user.
Agents govern themselves — the one unremovable safeguard is an out-of-band owner recovery (boot invariant + SSH script) the platform's own code can never disable. Locked out is impossible.
The owner is pinged only for results, human-only actions, and real emergencies — a daily briefing, an Owner Actions page (/action/:id) with reply/resolve, agent email (send + inbound on kasmo.dev), and session-limit auto-retry underneath.
Node + TypeScript · Fastify · SQLite · a React SPA · the Claude Agent SDK. One service serves REST + WebSocket and drives the agents.
Three under pm2: the app, the Discord bot, and the LiteLLM proxy (fallback engine) — behind Caddy on HTTPS.
Private GitHub mirror auto-pushed on every deploy — manual and autonomous. Runtime state lives outside the repo; secrets never leave the box.