Cone of Silence

Cone of Silence is our agent-pod service: your own always-on AI agent - Claude, working for you around the clock - running inside a hardened sandbox on a dedicated pod, reachable from your phone over Telegram. Message it like a person; it reads, researches, codes, and runs real tools - while everything it touches stays inside a security boundary we built and operate.

Always-on agents are the most exciting - and most exposed - thing happening in computing right now. Internet-wide scans this year found tens of thousands of personal AI agents running openly exposed, unsandboxed, with full access to their owners’ machines and credentials. The ecosystem’s defaults are convenience-first. Ours are not.

Security is the product

Each agent pod is built on Claudette, our open-source hardened agent sandbox, plus the same layered controls we deploy in our AI security engagements:

  • Sandboxed by default. The agent runs in OS-level isolation (bubblewrap) - it sees its project workspace, not the system. No “full access” mode, no opt-in security.
  • Scoped credentials. Per-project keys generated inside the sandbox; one project’s tokens can never leak into another’s. Your personal credentials never enter the pod.
  • Prompt-injection scanning. Everything the agent reads - web pages, files, command output, messages - is screened for injection patterns before it reaches the model.
  • Paired-device Telegram access. Your pod answers you, not whoever finds the bot. Pairing is explicit, allowlisted, and fail-closed - and access changes are never approved over the channel itself.
  • Egress controlled at the host, and an audit trail. The sandbox does not yet carry its own network namespace, so outbound traffic is governed by the pod’s host firewall and by the agent’s per-host network prompts rather than by the sandbox itself. Sandbox-level default-deny egress is on the roadmap, and we say so here because the mechanism is open source and you would find out anyway.

And unlike the original cone, ours actually works - the mechanism is open source, so you can check.

Managed, not abandoned

We provision the pod, harden it, pair your devices, and keep it patched and monitored. You get a monthly security review of what your agent did, what it touched, and what we tightened. When you outgrow one pod, we scale the same model to a fleet.

An honest threat model

Like Claudette itself, a pod defends an honest agent against hostile inputs - injection, credential theft, runaway tool use. It is not a magic box that makes AI output trustworthy, and we won’t pretend otherwise. Knowing exactly what a control does and doesn’t cover is the whole job.

Engagement & pricing

  • Pod onboarding - $1,500 one-time. Provisioning, hardening, Telegram pairing, credential scoping, and a working-session handover. Your scoping call is credited toward it.
  • Managed agent pod - from $195/month. Dedicated hardened pod, monitoring, patching, and the monthly security review. Bring your own Anthropic API key - your usage stays on your account, at cost.
  • Fleets & teams - by proposal.

Early-access pricing for design partners; indicative, finalized at scoping.

Start onboarding - $1,500 Book a scoping call first Create your account Questions? Ask us