Why it mattersA team can now run Claude Code and Codex as persistent workers on their own server, on the subscriptions they already pay for, without giving a third party access to the code or the credentials.
OtoDock is an open-source platform that runs Claude Code and Codex as long-lived agents on a server the user owns. The maintainer, Dimitris Mourtzis, posted version 1.6.0 to Show HN on 9 September, where it collected 46 points and 12 comments in 21 hours. The GitHub repository is at 117 stars.
Each agent is a persistent process running Claude Code or Codex against the user's own Anthropic or OpenAI subscription, or a local model. Agents keep their own memory, workspace, skills and tool set, and can be assigned to departments that delegate work to each other. The platform is multi-tenant by design, with four sharing modes for each agent, ranging from a private workspace per person to one shared history for the whole team.
What is in the box
The distribution ships the dashboard, an MCP tool catalog with GitHub, Notion, Home Assistant, Nextcloud, Prometheus, UniFi and Uptime Kuma agents pre-built, a speech stack, and a telephony daemon that connects agents to Twilio or a self-hosted Asterisk or FreePBX server so they can answer and place phone calls. Agents edit Word, Excel and PowerPoint files through a Collabora integration inside the chat, and the Claude Code terminal opens directly in the dashboard with the agent's tools already loaded.
A separate satellite process pairs any Mac, Linux or Windows machine to the server through one outbound WebSocket, so the same agent can act on the paired machine with no open inbound ports. If the machine goes offline, work falls back to the server.
Sandboxing
Server-side agents run inside a bubblewrap kernel sandbox with per-session mount and process namespaces. Network isolation is on by default through pasta: private IP ranges, the local LAN and cloud metadata endpoints are unreachable unless the admin grants an MCP tool scoped access. Credentials are encrypted at rest and injected only per session, so agents can use them without seeing them. SSO, two-factor auth and per-user cost budgets ship from the first install.
Cost and licensing
OtoDock is Fair Source under FSL-1.1-Apache-2.0, which converts each version to plain Apache 2.0 two years after release. Self-hosting is free for up to five users. Larger teams pay per seat.
The install is a Docker Compose stack that runs on 4 GB of RAM. Setup is one script that writes a .env, downloads a release-pinned compose file, and starts the platform on port 8400. Users connect their Claude or ChatGPT subscription inside the dashboard.
The trade the design makes is that all the agent state, all the credentials and all the traffic stay on hardware the user controls, at the cost of running that hardware. Nothing about the AI subscription changes: usage still counts against the same weekly and five-hour Anthropic caps the plans already carry.
Mourtzis says in the README that large parts of OtoDock were built by agents running on OtoDock itself, using Claude Code.
Source
Source: OtoDock
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually shipping with them. Short, and only when there is something worth reading.