5dive

5dive is a single Bash CLI (5dive, MIT-licensed) that provisions each agent as its own Linux user running an official coding-CLI binary (Claude Code, Codex, Antigravity, Grok, opencode, Devin, or the third-party hermes/openclaw harnesses) as a systemd service. Agents talk to each other by calling the same CLI — there's no message broker, no separate orchestration protocol, and the shared task queue is plain SQLite. The same binary that ships in the open-source repo is what runs 5dive's own managed VM product, so there's no open-core split between what you can self-host and what they sell. Requirements are a Linux host with systemd and root for install (or Docker if you lack either); a single-agent setup is advertised to run on a 1GB VM.

Each agent gets one of three isolation tiers set at creation: sandboxed (no sudo, own home, resource limits), standard (the default — shared read access but zero sudo), and admin (a narrow sudoers grant scoped to the 5dive binary itself, not NOPASSWD:ALL — it explicitly cannot run systemctl restart, edit a unit file, or sudo -u another user, because those are one-line root escapes). A fourth tier, beyond-admin, is full root and must be granted explicitly by an operator after the fact, never at creation. For git access specifically, 5dive push avoids handing an agent a long-lived credential at all: it holds one GitHub App key server-side, mints a short-lived repo-scoped token only after a human-cleared gate verifies, and discards it after the push — refusing pushes to main/master and to any branch without a cleared gate.

5dive runs against your own subscriptions or API keys — Anthropic Claude Pro/Max, OpenAI, Google, xAI, Cognition Devin, or OpenRouter/DeepSeek/Moonshot/Qwen/Z.ai via the claude type's --provider flag — with no 5dive account or OAuth proxy in between (per the vendor's "your auth tokens don't touch us" claim, which this review did not independently verify). One concrete, checkable constraint: as of April 4, 2026, Anthropic no longer permits routing consumer Claude Pro/Max OAuth through third-party harnesses, which is why the hermes and openclaw agent types (community multi-provider harnesses) require their own API keys rather than a Claude subscription — the vendor's own docs flag this directly rather than burying it.

A separate plugin (5dive-ai/5dive-council) adds a quorum-voting layer for agent decisions: seats vote, a chair frames the motion and votes last to prevent ties, the vote threshold scales with stakes (majority for routine calls, 2/3 with full attendance to amend the rules themselves), money/public-comms/secrets/destructive actions are human-only gates regardless of vote, and every result is sealed into a hash-chained, re-verifiable receipt log. A single human-held veto can stop any passed motion. This is opt-in — the base product's human-approval gates (tap-to-answer Telegram buttons) work without it.

The managed-VM pricing page (5dive.ai/pricing) redirects to a sign-in wall and returned a 403 to this review; the only pricing signal on the public site is "3 days free trial, cancel anytime" with no dollar figure anywhere reachable without an account. Self-hosting the open-source CLI is free. The company behind it, per its own Terms of Service, is TD The Market Publishers Ltd., a Cyprus-registered private company (HE186602, Limassol) — not a US or EU-mainland entity, which matters for anyone doing data-residency or vendor-due-diligence checks. The Terms cap 5dive's total liability at the greater of €100 or fees paid in the prior 12 months, which is a standard SaaS clause but worth knowing before running agents with server/sudo access against anything you can't afford to lose. SOC 2 is described on the Enterprise page as "on our roadmap" (not yet certified), and a DPA is available on request rather than published.

The 5dive-ai/5dive repository was created 2026-05-15 and had 64 stars, 11 forks, and 12 open issues as of 2026-10-04 (GitHub API). It is not archived and was pushed to the same day this was checked; releases are cut automatically whenever main is green (v0.71.1 shipped 2026-10-04). Unusually, the team's own claim that 5dive builds itself with its own agents is independently checkable on GitHub: merged PRs on the repo are authored by the 5dive-ai bot account with "Generated with Claude Code" in the commit body and reviewed/merged within minutes (e.g. PR #1230, opened and merged same day). A companion metrics page (5dive.ai/team) publishes daily shipped-task and human-ask counts to a public git branch — this is self-reported by 5dive's own measurement script, not independently audited, but the underlying mechanism (append-only git history) makes it at least tamper-evident rather than a bare marketing number.

5dive separately maintains @5dive/openagent (npm, latest 0.37.0; MIT), a standard for pinning an AI agent's name/face/voice/writing-rules in one YAML file and rendering it as an animated "trading card" with a gamified rarity tier (Common through Mythical) based on how complete the persona is. It's a genuinely separate, usable open-source tool, not vendor lock-in, but the rarity/collectible framing is unusual for a developer tool and worth knowing is adjacent marketing rather than core to running an agent team.