🛡

Hardened agent containers

Non-root user, setuid binaries stripped, minimal Linux capabilities, and configurable limits that default to 1 CPU / 8 GB / 3072 PIDs per agent container.

🌐

Proxy-enforced egress

iptables forbids direct internet. Squid is the only way out — with live-swappable whitelist modes (Strict / Default / Open) and per-domain access logs.

🔐

Zero default secrets

Web UI password, database password, and gateway key are generated randomly on first run. No "admin/admin" ever.

🖥

Live terminals in the browser

xterm.js over WebSocket, attached to the tmux session the agent is actually running in. Pause, watch, take over, reattach.

🧩

Multi-agent, one UI

Start with Codex, then add Claude Code, Gemini, Copilot, OpenCode, Pi, Cline, Continue, Grok Build, LLM, DeepAgents, Browser Use, OpenHands, or Bash through local installs and plugins. Same workflow, different tools.

💸

Cost tracker

Usage dashboards track Claude windows plus persisted per-agent token and cost estimates. Budget caps, exports, per-user breakdowns, and anomaly alerts are available as gated controls.

🛸

Context teleport

Hand off a stalled or rate-limited job to a fresh agent — git state, transcript digest, and task context packaged into a bootstrap prompt automatically.

🔁

Cron and queue automation

Schedule recurring or one-time jobs for Claude, Codex, OpenCode, and Pi, review execution history, and keep longer work organized through the job queue.

🔎

Browser verification

ShowOff drives a real browser to visually verify the agent's work — Playwright, DinD, or the Steel collaborative browser with agent-to-human handoff.

📋

Task board with dependencies

Persistent task state (pending → in_progress → to_review → completed), blocker graphs, and multi-assignee claim — so agents can resume work across sessions.

🗓

Cron scheduling

Run an agent nightly. Or every 10 minutes. Or once, at 3pm on Thursday. Schedule from the UI, manage routines centrally.

📝

Prompt templates

Save prompts with {{variable}} expansion. Share them across the team so everyone's agents get briefed consistently.

🖼

Image attachments

Drag-drop screenshots, mockups, or diagrams into prompts for multimodal tasks.

🧭

Git integration

Diff viewer, commit UI, worktree management, auto-sync. Review the agent's work like you would a PR — because it basically is one.

📡

Any OpenAI-compatible endpoint

LiteLLM, OpenRouter, vLLM, your own gateway. If it speaks OpenAI, Starbase can route to it.

🌏

i18n out of the box

15 locales: English, Chinese, Chinese (Hong Kong), Chinese (Taiwan), Dutch, French, German, Italian, Japanese, Korean, Polish, Portuguese (Brazil), Russian, Spanish, and Turkish.

📜

Structured audit logs

Every proxy request, every job transition, every agent action — written to the host filesystem, not stuck inside a container.

🔧

Manager mode

Let agents spawn and supervise other agents — safely — with the bin/sandbox orchestration CLI.

🚀

One-command deploy

make deploy HOST=user@host updates an already-cloned server with git fetch/reset, deployment metadata, and the documented production start flow.

See it running in five minutes.

Core self-hosting is free and Apache 2.0. Start with the documented open-source build path.

Read the quick start →