Agentic Vps

agentic-vps — provision & harden a box for autonomous agentic development

The verb for "make me another machine like the one we just set up." Distilled from the srv1692698 session (BRO-1550). It does not reimplement server hardening — it encodes the capability-preserving model and the lockout-safe staging sequence so a future agent reaches for it instead of re-deriving it.

The one invariant

The box IS the sandbox. Security comes from containing the blast radius (non-root agent user · VPN-only access · snapshot rollback · no long-lived secrets in the agent's reach), not from restricting the agent (no tight egress allowlist, no permission prompts, no seccomp leashes). The agent gets the same freedom it has on a local machine — full sudo, open network, any dep, YOLO mode — because a compromise is contained and reversible, not prevented.

This is the user's standing constraint: "secure but don't limit capability/autonomy." The model and its threat-model rationale live in research/entities/concept/lethal-trifecta-denial.md and the playbook at docs/security/2026-06-24-vps-agentic-hardening-playbook.html.

Hard rules (the lockout-safety sequence)

These are non-negotiable ordering constraints. scripts/staging_check.py enforces them deterministically — a plan that violates them fails the gate.

1. Snapshot before any risky change. A rollback point must exist first.
2. VPN up + verified BEFORE closing public SSH. Bring up the mesh VPN (Tailscale/WireGuard), then in a second, independent session confirm ssh agent@<vpn-ip> works — only then remove the public :22 rule.
3. Allow the new path before removing the old. UFW: allow in on <vpn-if> before delete allow OpenSSH. Edge firewall: add the VPN rule, verify, then drop TCP/22 and sync.
4. Verify on a fresh connection after every SSH/firewall change — never assume the change is safe from the session that made it.
5. Out-of-band recovery must exist (cloud Browser Terminal / recovery mode) before public SSH is closed, since there is no SSH fallback afterward.

Procedure

Phase 0 — pre-flight (latent: per-provider control plane)

• Identify the provider control plane (Hostinger VPS MCP, AWS, etc.).
• Snapshot the box (Hostinger: VPS_createSnapshotV1).
• Confirm out-of-band console access exists.

Phase 1 — provision (deterministic: scripts/provision.sh)

Run as root over SSH — idempotent, safe to re-run:

ssh root@<host> 'AGENT_USER=agent bash -s' < scripts/provision.sh

Creates the non-root agent user (passwordless sudo + docker), installs the toolchain (Node, Bun, Rust, gh, Docker, Claude Code + sandbox-runtime), generous resource caps (host-survival, not throttle), a bypassPermissions Claude config, and git identity. Tailscale is installed but auth is interactive (prints a login URL).

The config also sets denyRead on credential dirs, but that only takes effect if you flip sandbox.enabled=true — in the default capability- preserving (sandbox-off) mode it is inert. The primary secrets protection is no long-lived secret in the agent's env (Phase 2: interactive OAuth, scoped tokens), which verify.py's no_secrets_in_env gate actually checks.

Phase 2 — credentials (latent: user authenticates interactively)

The user SSHes in and logs into the CLIs themselves — never paste long-lived broad keys into env:

• claude → /login (OAuth, stored in ~/.claude/.credentials.json)
• gh auth login && gh auth setup-git (prefer a fine-grained PAT scoped to specific repos over a broad repo OAuth token)

Phase 3 — perimeter (deterministic order; see Hard rules)

• tailscale up --hostname=<name> → user authorizes the printed URL.
• Verify ssh agent@<tailnet-ip> in a fresh session.
• UFW: allow in on tailscale0 + allow 41641/udp, then delete allow OpenSSH.
• Edge firewall: add UDP 41641, remove TCP/22, sync.

Phase 4 — verify (deterministic: scripts/verify.py)

Collect facts from the box and assert the capability-preserving invariants:

python3 scripts/verify.py --host agent@<tailnet-ip> [--public-ip <ip>]

Gates (required): non-root agent user · resource caps set · Tailscale up · public :22 closed · no long-lived secret in agent env. Recommended: snapshot exists · sandbox config present · git identity set.

What's deterministic vs latent

Anti-rationalization

References

• research/entities/concept/lethal-trifecta-denial.md — the security model
• docs/security/2026-06-24-vps-agentic-hardening-playbook.html — full rationale
• docs/security/srv1692698-runbook.html — operations (access/recover/rebuild)
• docs/security/srv1692698-provision.sh — the original instance this generalizes