LiteLLM-Labs just dropped Lite-Harness on Hacker News, and if you're running multiple AI coding agents in your stack, this might be the glue you've been missing. It's a self-hosted harness server that lets you deploy agents from Claude Code, Codex, OpenCode, and GitHub Copilot to a unified backend—complete with durable sessions, cron scheduling, vault storage for credentials, and sandbox isolation out of the box. The whole thing ships as a single Dockerfile, which is exactly the kind of minimal-friction deployment that gets ops teams excited.

What You're Getting Out of the Box

The pitch is straightforward: stop maintaining separate services for each agent harness. Lite-Harness wraps everything in an OpenCode-compatible server so you get shared MCP tools, unified system prompts, and consistent session management across all your harnesses. Agents become accessible via CLI, API (following the Open-code API Spec), and even Slack integration. The vault feature handles credential storage—think BROWSER_USE_API_KEY or LINKEDIN_PROFILE_ID—so you're not hardcoding secrets into agent configs. Cron scheduling works too: deploy an agent that runs every four hours on weekdays with human approval gates before each execution, all visible through the Inbox UI.

Deployment Is Dead Simple

Getting started requires only a few environment variables and Docker. You generate a master key with openssl rand -hex 32, point Lite-Harness at your existing LiteLLM gateway (LITELLM_API_BASE and LITELLM_API_KEY), and you're rolling. The web UI lives on port 4096 by default—paste your master key and you're in. For persistence, mounting ./data:/home/.local/share/lite-harness preserves sessions across restarts, or you can set a custom DB_PATH for cloud volume mounts like /mnt/data/db.db. On restart, the server logs rehydrated N sessions from the database and everything's immediately available again.

Sandboxing With E2B or Daytona

Security-conscious teams will appreciate that setting an E2B_API_KEY or DAYTONA_API_KEY environment variable automatically spins agents into isolated Linux sandboxes. This means your LinkedIn outreach agent, your stargazer DM bot, whatever you're running—it's not touching your host system. Templates and snapshots are supported for more advanced sandbox configurations. The docs point to full configuration options including vault setup and custom harness templates.

Why They Built It (And the Real Pain Point)

The about section cuts right to it: running opencode and claude-code as separate servers became unmaintainable. Multiple services meant different API specs, unreliable session management, divergent inputs for MCP tools and system prompts. Lite-Harness solves this by consolidating everything into one service you maintain—one config file, one deployment, one set of credentials flowing through the vault. It's a classic infrastructure consolidation play that resonates if you've ever debugged why your Cursor agent works but your Claude Code instance is silently failing on the same task.

Key Takeaways

  • Single Docker container unifies Claude Code, OpenCode, Codex, and GitHub Copilot harnesses under one API
  • Durable sessions survive server restarts via mounted storage or custom DB_PATH configuration
  • Vault system securely stores API keys and credentials without hardcoding them into agent configs
  • Cron scheduling with human-in-the-loop approval available through the Inbox UI
  • E2B and Daytona sandbox integration keeps agents isolated from your host infrastructure
  • MIT licensed, fully open-source on GitHub at LiteLLM-Labs/lite-harness