Autonomous agents execute infra commands — one unchecked terraform destroy and your production database is gone. SoloShip wraps every command in a plan: danger-classified, guardrail-checked, AI risk-scored — and critical plans cannot execute without your explicit approval.
No agent framework. No lock-in. Every execution sealed into a tamper-proof vault.
Paste the command your agent wants to run — rm -rf /data, terraform destroy, systemctl status nginx — plus the target environment and agent name.
SoloShip classifies danger — safe, destructive, critical — checks region pinning and backup verification, generates a rollback plan, and scores the risk with AI (deterministic fallback if the AI is down).
Critical plans are blocked until you approve. Executions run through the gated wrapper and every one is hash-sealed into the evidence vault — a receipt no one can rewrite.
Deterministic regex rules sort every command into safe, destructive, or critical — terraform destroy, db wipes, volume deletes, region migrations, and prod+delete combos never slip through.
Region pinning keeps staging workflows away from prod. Destructive plans require a verified backup. Every plan ships with a generated rollback plan.
GateAI scores blast radius 0–100 with concrete factors. If the AI gateway is unreachable, a deterministic heuristic answers instead — same shape, labeled fallback, never a 500.
Critical plans return 403 without approved=true. Every execution is sealed into the hash-chained evidence vault with a full audit trail, newest first.
For the first projects you dare to automate.
For founders running real deploy agents.
For operators running fleets of named agents.
Not in v1. The executor is simulated: it records the gated execution — exit code, output log, applied-at timestamp — and seals the record into the evidence vault. The classification, guardrails, approval gate, and audit trail are all real. Real host adapters arrive after validation; the guardrails won't change when they do.
Deterministic regex rules, checked in order. Critical patterns first — terraform destroy, database wipes, volume deletes, region migrations, and any destructive command targeting prod. Then destructive patterns — rm -rf, drop, delete, resize, migrate, force flags. Everything else is safe. Same input, same answer — no black box.
Nothing breaks. GateAI falls back to a deterministic heuristic that scores the plan from its danger class, environment, region and backup posture. The response is labeled mode: "fallback" so you always know which path produced the score. AI output is advisory — it never blocks or auto-approves a plan.
No. A critical plan returns HTTP 403 from the execution endpoint unless the request carries approved=true. There is no override flag, no API key, no agent-side bypass. Approval is a human action by design.
Gate the next command in under a minute — free, with a sealed audit trail from day one.
Open the app — free