OtoDock: Self-Hosted Multi-Tenant Platform for Claude Code and Codex Agents
OtoDock is a self-hosted platform that runs multiple AI coding agents — powered by Claude Code or Codex CLI — as persistent processes on your server. It coordinates them into what the author calls a "company OS": agents get departments, schedules, shared or private workspaces, and even phone numbers. Today's release is version 1.6.0, and everything is public under a fair-source license; self-hosting is free for up to five users.
Core architecture and sandboxing
Each agent runs on your server in a kernel sandbox (bubblewrap) with network isolation always on (pasta). Every agent gets its own workspace, memory, and schedules, and can be configured to run identically on any remote computer using a single outbound WebSocket — no inbound ports or VPN needed. You can bring your own Anthropic subscription (for Claude Code) or OpenAI subscription (for Codex), or run local models. The engine is chosen per agent and can even be switched per chat.
Agent anatomy and sharing modes
Each agent has six editable components — persona, memory, workspace, knowledge, skills, and tools. Four sharing modes control how work lands:
- Personal only: a private workspace for each person.
- Personal + shared: private default, plus a team workspace.
- Shared + personal: shared default, plus a private space.
- Shared only: one history and workspace for the whole team.
Roles are tiered: Admins run the platform, Managers control agents, Creators make them, Editors modify shared files, Members and Viewers use or read only. Multi-tenancy means many people can collaborate on the same agents without stepping on each other's work.
Built-in features
Agents aren't limited to coding. They can make and receive phone calls through Twilio or your own Asterisk server, edit and preview videos, Excel, Word, and PPT files inline via Collabora, and notify you with four severity levels. They can run on schedules, on webhook events, or on demand. Each run is a full chat that can be reopened and continued — you don't have to watch it live.
The author notes that large parts of OtoDock were written by OtoDock itself running Claude Code, and includes this challenge for HN: "I would love people to read the code, take a look in the sandbox model, and tell me what would stop you from running this on your own hardware."
📖 Read the full source: HN AI Agents
👀 See Also

Time Complexity MCP: Static Analysis Tool Feeds Big-O Complexity to AI Coding Agents
Time Complexity MCP is an open-source MCP server that performs static code analysis to detect Big-O complexity, feeding the results directly to AI coding agents like Claude Code or Copilot without token consumption. It supports JavaScript, TypeScript, Python, Java, Kotlin, and Dart.

Agent Factory: Autonomous System Builds AI Agents from Online Problem Discussions
Agent Factory is an autonomous system that scrapes Reddit, HN, GitHub, and Twitter for real problems, scores them on demand, market gap, and feasibility, then builds standalone AI agents for promising ideas. The system uses a minimal Next.js template with 7 tools and runs Claude Code headless via a shell script.

Unsloth and NVIDIA Collaborate to Speed Up LLM Training by ~25%
Unsloth and NVIDIA release optimizations for LLM training: caching packed-sequence metadata (~14.3% speedup) and double-buffered async gradient checkpointing (~8% speedup), with no accuracy loss. Auto-enabled on RTX laptops, data center GPUs, and DGX Spark.

Building a Programming Language with Claude Code: The Cutlet Experiment
Ankur Sethi built a complete programming language called Cutlet using Claude Code over four weeks, with the AI generating every line of code while he focused on guardrails and testing. The language features dynamic typing, vectorized operations, and a REPL, running on macOS and Linux.