5dive is an open-source system for running a team of AI agents on your own server, with the core idea of "using the operating system instead of building your own platform." Each agent is a separate Linux user running an official coding CLI (such as claude, codex, antigravity, grok, openclaw, hermes, opencode, pi, etc.) as a systemd service, coordinating through a shared bash CLI. 5dive itself introduces no frameworks, protocols, or message brokers, relying only on Bash, SQLite, and systemd to accomplish orchestration, so its resource footprint is extremely low, and a single-agent configuration can run on a VM with 1GB of memory.
In terms of use cases, 5dive is suited for users who want multiple AI agents to collaborate autonomously and take on real workflows. Agents claim tasks from a shared SQLite task queue, can talk to each other, and hand off work while the user is asleep. When something arises that only a human can decide (such as spending, publishing, or destructive operations), the task is suspended and a notification with buttons is sent via Telegram; after the user clicks a button, the agent that owns the task can unblock it and continue. Each agent can have its own channel (such as Telegram), or the whole team can share a single bot, creating a separate forum topic in the group for each agent.
Main capabilities include: creating a self-organizing company structure with one command via 5dive company; 5dive task manages the shared task queue and handoffs between agents; 5dive goal turns objectives into a protected task graph; 5dive loop lets agents perform bounded autonomous work in persistent loops; 5dive council provides optional audit-style adversarial review; 5dive trace and 5dive run provide a complete causal timeline of a task and receipts and metrics for each attempt; 5dive watch and 5dive wall monitor the team in real time; 5dive project supports one team running multiple products; 5dive org maintains the reporting relationship graph; 5dive trigger turns signed GitHub or generic webhook events into ordinary tasks; 5dive memory provides persistent memory search with provenance; 5dive acp supports connecting clients such as Zed and Buzz via ACP. In addition, the official offering also provides 5divThe e-ui plugin is used for a local browser view, and the 5dive-browser plugin provides a shared, human-authenticated browser for the entire team.
In terms of usage, users need a Linux host with systemd, as well as their own agent CLI subscription or API key (such as Claude Pro/Max, OpenAI, etc.), and do not need to register an account with 5dive. Quick installation is performed via curl | sudo bash; the installation script only installs dependencies via apt and places the CLI and systemd units. When each agent is created, three isolation levels can be chosen: the admin level is not root, and is only granted passwordless permission for all 5dive subcommands, and cannot restart services or edit unit files; scenarios requiring the entire host should be authorized in a declarative way. Users can also configure a shared authentication profile via 5dive account set, which supports Anthropic-compatible endpoints (--base-url), and can point a shared bot to a Telegram group. 5dive supports importing an entire organization template (5dive team ls, 5dive up with 5dive.yaml) or a character pack to quickly assemble a team, and plugins can add channels, verbs, or capabilities to all seats at once. Plugins are divided into two categories: official (from github.com/5dive-ai) and community. Before installing a community plugin, the repository and commit information are displayed and a warning is given that it runs without a sandbox.
Applicable scenarios include: building an autonomously operating AI agent team on your own server, running multi-product projects, automated workflows that require humans to intervene at key decision points, remotely managing agents via Telegram, and wanting to use your own API keys without relying on a third-party intermediary.
In terms of limitations: 5dive requires users to provide their own Linux host and AI CLI subscription or key; isolation depends on Linux users and systemd, and non-Linux environments are not yet supported; plugins have no remote kill switch or signature, and community plugins require users to evaluate trust on their own; as of April 4, 2026, Anthropic no longer allows routing consumer Claude Pro/Max OAuth through third-party harnesses, and such scenarios must use the official claude type and provide their own API key; --base-url requires HTTPS (except for localhost), and must be used with an authentication profile. Other information regarding pricing, hosted VM details, platform support scope, etc., is not yet available.
