Orva

Self-hosted Functions-as-a-Service for your homelab or on-prem server.

Write a JavaScript, TypeScript, or Python function, hit deploy — Orva runs it in an isolated nsjail sandbox and serves it over HTTP. One Docker container gives you the runtime, a dashboard, a CLI, an MCP server, and a built-in AI assistant. It's for the Lambda/Workers workflow — write a function, invoke it over HTTP, schedule it, chain it — on hardware you control (a Pi, a homelab box, a VPS, bare metal). No cloud account, no per-invocation billing.

Active development. Solid for homelabs, side-projects, and internal tools. Not recommended for customer-facing production yet.

• Two runtimes — node (Node.js 24, also runs TypeScript) and python (Python 3.14).
• Real isolation — every function runs in its own nsjail sandbox: user namespace, chroot, cgroup v2 limits, and a seccomp syscall allowlist. → Security
• Warm pools — idle workers stay resident per function, so repeat calls skip cold start.
• Built-in primitives — per-function KV store, background jobs (retries + backoff), cron schedules, function-to-function calls, encrypted secrets, custom routes, and signed inbound webhooks.
• Distributed tracing — every HTTP → F2F → job chain shares one trace, with a waterfall view and zero code changes.
• Versioning — content-hashed deploys with one-click (or one-command) rollback and side-by-side diffs.
• MCP + AI — a 73-tool operator MCP server at /mcp and a built-in agentic AI assistant (dashboard or orva chat) that operate your instance with your own provider key. → AI & MCP
• Templates — 21 starters (Stripe/GitHub webhooks, JWT/OAuth, CSV→JSON, URL shortener, …) in the editor.

Docker Compose (recommended for persistent setups):

curl -fsSL https://raw.githubusercontent.com/Harsh-2002/Orva/main/docker-compose.yml -o docker-compose.yml
docker compose up -d

Compose publishes on http://localhost:3000 (it maps 3000:8443), not :8443. Open it and finish the short onboarding flow.

Bare metal / VM (systemd or OpenRC):

curl -fsSL https://github.com/Harsh-2002/Orva/releases/latest/download/install.sh | sh

The installer supports Debian/Ubuntu, Fedora/RHEL/Rocky/Alma, Alpine, Arch, and openSUSE. It verifies a real sandbox as the unprivileged orva service user and stops with a host-specific diagnosis if the kernel cannot run Orva securely.

CLI only (operator laptop or CI runner):

curl -fsSL https://github.com/Harsh-2002/Orva/releases/latest/download/install-cli.sh | sh   # macOS / Linux
irm  https://github.com/Harsh-2002/Orva/releases/latest/download/install-cli.ps1 | iex        # Windows

Installers are idempotent — re-run to upgrade; pin a version with ORVA_VERSION=vYYYY.MM.DD. TLS, reverse proxy, and backup guidance: docs/DEPLOYMENT.md.

After onboarding, connect the CLI and deploy:

orva login http://localhost:3000
orva deploy ./src --name my-fn --runtime node   # runtimes: node | python
orva invoke my-fn --body '{"name":"world"}'
orva logs my-fn --follow

CLI reference: docs/CLI.md.

Defaults work out of the box. These are all supported server environment variables:

See docs/CONFIG.md for validation rules, security implications, and examples.

The orva SDK is preinstalled in every sandbox — KV, function-to-function invoke, and background jobs, with nothing to install:

// Node — Python uses the same shape: from orva import kv, invoke, jobs
const { kv, invoke, jobs } = require('orva')

exports.handler = async (event) => {
  await kv.put('hits', (await kv.get('hits') || 0) + 1)
  await invoke('send-notification', { msg: 'hello' })   // child span in the same trace
  await jobs.enqueue('audit-log', { at: Date.now() })   // async, retried on failure
  return { statusCode: 200, body: { ok: true } }
}

The SDK reaches Orva over HTTP, and network_mode defaults to none — so set the function's network mode to egress or every SDK call fails with ENETUNREACH. The editor warns you at deploy time when it spots the import.

Handler contract, event shape, and streaming: docs/RUNTIMES.md.

The dashboard's AI section and orva chat operate the instance using your own OpenAI, Anthropic, or OpenAI-compatible provider key, with optional approval before writes.

To use an external MCP client such as Claude Code or Cursor, add:

https://your-orva-instance/mcp

The MCP server can create and deploy functions, invoke them, read logs, manage secrets, and browse KV. See the Orva reference.

Apache-2.0