Krowoc

Your agent, in the cloud. A web app (PWA, no install) that runs an AI agent with a real shell, files, browsing, and connectors — 24/7, on its own cloud machine. Open it from any device; close your laptop and the task keeps running. Open source, bring your own keys.

• 🔌 Engine-agnostic — Claude Agent SDK (your Anthropic API key), open models via Fireworks/OpenRouter/Ollama, or Eigent's multi-agent workforce
• ⚡ Scheduled triggers — "every morning at 8, brief me" actually runs
• 🌐 Real browsing — headless Chrome via Playwright MCP
• 🧩 MCP connectors — attach Composio/Klavis/Zapier MCP or any endpoint
• 🔑 BYO everything — paste a key, pick a model; keys never leave your box

Scheduled triggers with run history	Engines, connectors & BYO models

phone / laptop browser
   → PWA + control plane        (web/, control-plane/ — zero-dep Node)
      → agent engine            (per user, sandboxed)
          eigent   — multi-agent + browser automation   [primary]
          opencode — lightweight coding agent           [fallback]
      → LiteLLM proxy           (litellm/ — Fireworks default, BYO keys)
      → MCP connectors          (Composio/Pipedream to start)

Status

• Control plane: sessions + SSE streaming proxy, pluggable engines
• PWA UI: chat with history sidebar, markdown rendering, scheduled-tasks panel, installable, dark/light
• OpenCode engine — live-tested with Kimi K2.6 on Fireworks (real tool use, streaming)
• Claude Agent SDK engine (engine-claude/) — live-tested; runs on the user's own Anthropic API key (the sanctioned third-party path)
• Bearer-token auth (KROWOC_AUTH_TOKEN) — required for public deploys
• Scheduled tasks: daily/interval prompts fire in fresh agent sessions
• Eigent engine adapter + browsing Dockerfile (written; container untested)
• Cloud deploy (single box via docker-compose, then Fly Machines per user)
• Supabase auth (multi-user), connectors panel (Composio MCP), web push

Host it on Fly.io (recommended, ~10 min)

One command sequence gets you a personal cloud agent with HTTPS, phone PWA, browsing, and 24/7 scheduled tasks — no Docker needed locally. → docs/deploy-fly.md

Run locally (no Docker)

# 1. engine
npm i opencode-ai && npx opencode serve --port 4096
#    give it a provider config: see agent/opencode.json (uses LITELLM_URL)

# 2. models — either run LiteLLM (needs FIREWORKS_API_KEY):
pip install 'litellm[proxy]' && litellm --config litellm/config.yaml --port 4000
#    ...or any OpenAI-compatible endpoint

# 3. control plane + web UI
node control-plane/src/server.js     # http://localhost:8080

Run the full stack (Docker)

cp .env.example .env   # fill in FIREWORKS_API_KEY + LITELLM_MASTER_KEY
docker compose up --build            # OpenCode engine
# or the primary engine (heavy image, includes Chromium for browsing):
git clone --depth 1 https://github.com/eigent-ai/eigent vendor/eigent
ENGINE=eigent ENGINE_URL=http://agent-eigent:5001 \
  docker compose --profile eigent up --build

Caddy terminates TLS on $DOMAIN and proxies to the control plane.

Design decisions

• Per-user isolation = Fly Machines (microVM + persistent volume, scale-to-zero; idle user ≈ $0.50/mo, active ≈ $1–3/mo). Template in infra/fly/. Migration path to self-hosted Firecracker at scale.
• Engine keys never leave the platform: the Eigent adapter injects the LiteLLM route per request; engine containers hold no provider keys.

See docs/architecture.md for the full engine evaluation and roadmap, and SECURITY.md for credential handling and deployment hardening.

Community

• Discord — help, ideas, show-and-tell
• Contributions welcome — see CONTRIBUTING.md

License

Apache 2.0 (see LICENSE). The Krowoc name and logo are not covered by the code license. This public repo is the complete single-tenant product; the hosted multi-tenant cloud (accounts, provisioning, billing) is a separate closed layer built on top of it.