AgentHost | Your AI agents keep working. You stay in control.

Session-bound tools. Close the lid, the agent forgets everything. Memory, tools, work state — all gone.

No guardrails. One bad instruction can reach files, money, or production before you can stop it.

Someone else's backend. Your credentials, memory, and work history belong in your cloud account, not an agent vendor's.

Watch a merge get stopped.

Five agents review a pull request on one box. The red-team agent finds a concurrency hazard and fails it. A hotfix goes in, the council re-runs, and it merges — under the strictness level the operator set.

The council in this clip is scripted. The real one streams live from your box.

What changes<br>From session-bound to persistent.

Your agents keep their memory, tools, and work state across sessions — with budgets, approvals, and audit trails you control.

YOUR AGENT, INTACT<br>Four agent interfaces and one messaging gateway, on infrastructure you own.

Claude CodeCodexHermes AgentGemini CLIOpenClaw

Claude Code and Codex can execute and cross-review approved autonomous tasks. Hermes Agent and Gemini CLI join chat workflows. OpenClaw runs separately as the Telegram and Discord gateway. Each arrives with its identity intact — skills, memory, MCP servers, plugins — on the same box.

GOVERNED BY DESIGN<br>Give agents a small, controlled workspace.

Read-jailHuman gate6-run cap45-minute cap150K tokens$5 chain capCross-engine review

Spend caps per work chain. Human approval gates on deploy, delete, send, and credentials. Cross-engine review before work is marked done. Every action audited. Deploying, spending, deleting, email or messages, and ambiguous work wait for a human. Social publishing can run unattended only through the narrow Postiz allowlist.

The first production run found four bugs. The first live autonomy day surfaced four real bugs on the founder's production box — and the governance system caught every one.

The gates stopped work for review. Tasks stopped at the gates, budgets capped runs, and cross-engine review rejected bad results before they reached production.

The code and audit logs are public. The CLI is public on npm. The production log is the proof until real customers exist.

// the difference<br>The migration is the moat.

Anyone can start a cloud container. The hard part is turning a session-bound tool into a persistent, governed engineering system. That's what AgentHost does.

Watch one task move. The board keeps its owner, state, reviewer, and next decision visible. When the work needs you, it lands in Awaiting Your Review and the phone alert opens the exact card.

agenthost▮ Control RoomChatLoops

launch-page review · one taskawaiting operator

01Queued Claude assigned

02Working draft complete

03Independent review Codex checks raw result

04Awaiting you decision required

AWAITING YOUR REVIEW<br>Checkout disclosure needs your call.

Codex rejected the first pass and attached the reason. Your phone alert opens this exact task.

ApproveSend backReassignOpen

WORKS WHILE YOU'RE AWAYSchedule agents that keep working across sessions.<br>A 7am briefing reads your repos and memory and hands you a digest before you're up.<br>24/7 across sessions

TELLS YOU WHEN IT'S DONEPush notifications when a scheduled run or a chat finishes.<br>The box sends them itself — no vendor backend required.<br>1 tap opens the exact card

STANDING TEAM CHARTERThe engines enter every turn with the same rules.<br>Roles, handoff format, shared-board behavior, budgets, and blocked-work reporting load for Claude, Hermes, and Codex — persistent across sessions.<br>roleshandoffsreview

GOVERNED BY DEFAULTAutonomy is off by default, with a stop control always visible.<br>Eligible Claude and Codex tasks run in restricted environments. Risky or ambiguous work parks for you, and every chain has hard ceilings.<br>6 runs45 min150K tokens$5

FILES BOTH WAYSThe work can leave the chat.<br>Upload files to the box, download finished artifacts, or send text back to your desktop through the optional bridge.<br>inboxdownloadsdesktop bridge

// your agent, intact<br>Your agent arrives as itself, not as a blank container.

Skills, memory, MCP servers, plugins — all migrate. Keep Claude Code, Codex, Hermes Agent, Gemini CLI, OpenClaw, MCP, memory, and messaging on one box, with the Control Room above the work. AgentHost adds persistent infrastructure; supported tasks gain budgets, review, and a durable record.

AGENTS | GATEWAY | MEMORY | CHANNELSKeep the engines, tools, and channels you already use.

Real product marks identify supported tools and channels. They do not imply endorsement, sponsorship, or partnership.

STARTER STACKINSTALLED ON FIRST BOOT

01Engineering<br>Superpowers, AgentMemory, and CodeGraph for planning, testing, review, memory, and local code context.<br>included

02Growth<br>Claude Ads, Claude SEO, Claude Blog, and Marketing Skills for research through distribution.<br>included

03Launch<br>Newsjacking and Pitch Deck Mastery arrive as standalone skills you...