Get started

we open-sourced 5dive

the thing that runs our cloud is on github now. 5dive — one bash file, MIT-licensed, the same binary that manages every agent on every 5dive runtime vm.

you can install it the old way:

curl -fsSL https://install.5dive.com | sudo bash

but that’s not the install path we expect most people to use.

tell your agent to install it for you

if you already have claude code, codex, hermes, openclaw, or opencode open in a terminal, paste the prompt from the readme into it. the agent runs the installer, confirms 5dive --version, installs the 5dive-cli skill into its own runtime, and asks you which agent to create first.

then you keep talking to the same session. “create a research agent paired to telegram.” it does the right 5dive agent create invocation, prompts for the bot token, runs the pairing flow. no second tool. no separate dashboard tab. no --help rabbit hole.

if your agent is on a laptop and the target is a remote vm, the readme has an ssh addendum — prefix the install line with ssh -t <user@host> and the agent does the right thing.

the surprise: agents call it too

we built the cli for humans. then the agents found it.

every agent on a 5dive host is a linux user. give it sudo 5dive ... in its sudoers, and it can spawn, message, and tear down siblings on the same box.

# main → marketing: hand off a writing task (fire-and-forget)
5dive agent send marketing "draft a launch tweet for v0.4"

# marketing → main: pull a fact before drafting (wait for the reply)
5dive agent ask  main "summarize the v0.4 changelog — 3 bullets" --timeout=120

no coordinator service. no orchestration layer. no json-rpc gateway. the binary on the host is the bus.

spawn a senior and a junior, give them different prompts, let them pass work. read every line of what they say to each other in the same logs you read your own commands in.

what it is, briefly

5dive runs ai coding agents — claude code, codex, hermes, openclaw, opencode — as persistent processes on a linux host. each agent is one linux user, one tmux session, one systemd unit. you log in with your own credentials. we never proxy your auth. the binary doesn’t phone home.

5dive agent create scout --type=claude
5dive agent send scout "summarize today's logs"
5dive agent ask  scout "what's the status of the deploy?" --timeout=60
5dive agent scout tui   # attach to the live session

that’s the surface. the rest is variations — telegram channels, isolation tiers, skills, a compose verb for declaring a team in yaml.

what’s in v0.1.1

short list, lifted from the changelog:

  • agent install path: paste a prompt into your existing claude code / codex / hermes / openclaw / opencode session. the agent installs the binary and the skill that drives it.
  • lifecycle: agent create / list / start / stop / restart / send / ask / logs / tui / rm.
  • auth: 5dive auth set | login | status per type. shared profiles via 5dive account so a team of agents can share one login.
  • isolation tiers: admin, standard, sandboxed. give an untrusted agent its own home dir; give a trusted one full host access.
  • channels: telegram + discord. pair a bot, message your agent from your phone.
  • compose: declare a team in yaml, stand it up in one command.
  • doctor: 5dive doctor --repair --json — health check, auto-fixes, machine-readable for ci.
  • skills: prompt bundles that drop into any agent. browse the skills repo.
  • docker demo: kick the tires without touching your host.
  • json envelope: --json on any command. {ok:true,data} or {ok:false,error:{code,class,message}}. exit code matches error.code.

the whole installed cli is a single bash file at /usr/local/bin/5dive.

what it doesn’t do

a few things on purpose:

  • no remote control plane. you don’t point one cli at a fleet of vms from your laptop. each host is its own world. failure modes stay local.
  • no plugin system. the verbs are flat. we’ll add commands when we have to, not before.
  • no telemetry, ever. the cli does not phone home. it only talks to the model providers you configure.
  • linux only. macos and windows via the docker image (docker run --privileged 5dive). native macos is on the list but not soon.

want a dashboard?

the cli is the oss surface. if you’d rather click than ssh, 5dive.com is the managed version — same cli under the hood, but the vm, hardening, backups, allowlists, and the dashboard are run for you.

the cli is the same binary either way. that’s the contract.

what’s next

the agent-to-agent surface is the part most likely to grow. stateful conversations, typed handoffs, multi-turn negotiations — we want to design those in public. file an issue or open a pr if you have opinions.

repo: github.com/5dive-com/5dive skills: github.com/5dive-com/skills bugs and feature asks: github issues on either.

mit. one bash file. one prompt to install it. yours now.