The AI agent OS,
with the brakes on.
Everything OpenClaw and Hermes-Agent do — Telegram, Discord, voice, scheduled tasks, MCP, sub-agents — in a hosted dashboard with spending caps, approval queue, and live replay built in. No 15-hour setup. No $5k surprise bills.
Drafted a three-sentence reply that confirms the updated pricing and proposes a 15-minute call Thursday afternoon.
Open-source agents are powerful. They're also dangerous.
We took the parts that make OpenClaw and Hermes feel like AGI, then put them behind brakes that prevent the parts that make them risky.
Spending caps that actually stop spending
Set a monthly cap. The agent's checked before every LLM call, paused automatically at 100%, and you get alerts at 50/80/100. No $5k overnight horror stories.
Mission Control with full replay
Live event stream of every LLM call, tool execution, sub-agent spawn, and cost. Click into any past run for a step-by-step trace. The truth of what your agent actually did.
MCP ecosystem on tap
Plug in any Model Context Protocol server — GitHub, Stripe, Postgres, Notion, hundreds more. Your agent gets new tools the second you connect.
Approval queue, deferred execution
Risky actions like sending email don't run until you approve them in the dashboard. The result posts back to the conversation. Real safety, not theater.
One agent. Every channel.
Same memory, same skills, same Mission Control trace — whether the message comes from Telegram, Discord, the browser, or your own voice.
Telegram
Full bot via @BotFather. 5-minute setup.
Discord
Webhook posts + /zexo slash commands via Interactions API.
Voice
Browser mic → Whisper → agent → TTS playback.
Web chat
The dashboard is itself a channel.
Email-via-@zexo.ai address coming next. WhatsApp queued behind Meta Business API approval.
The full agent OS, without the bloat.
Six features the OpenClaw “5 features” analysis says you need. Plus the discipline to keep your agent's context clean — no 45k-token system prompts that destroy quality after a month of use.
Tiered semantic memory
Identity / essential / project / reference tiers with pgvector hybrid search. Your business context stays loaded, everything else is one tool call away.
Brand & identity layer
Tell your agent who you are, what you sell, and how you talk — once. Every run inherits that context, on every channel.
Scheduled tasks
Cron-pinned proactive runs. Your agent works while you sleep — morning briefings, weekly digests, Monday triage.
Multi-goal kanban
Track parallel goals like a product team would. Each goal has a plan the agent reads and updates as it works.
Curated skill library + your own
13 vetted skills out of the box: web search, email, notes, Discord, files, sub-agents, time, scheduler. Plus any MCP server.
Sub-agent orchestration
Your main agent can delegate focused tasks to specialist sub-agents. Cheaper models for narrow work, full traces in Mission Control.
Won't spend money. Won't send anything. Won't go rogue.
The OpenClaw subreddit is full of $5k overnight runs and accidental email blasts. Zexo's safety isn't bolted on — it's the runtime, the queue, and the kill switch checking every step.
Spending caps
Set a monthly ceiling per workspace. Your AI stops before the bill surprises you.
Deferred-execution approvals
Risky actions create an approval row instead of running. You decide; we then execute and post the result.
Kill switch in one click
Pause an agent or the whole workspace from Mission Control. Persists across the entire runtime — no race conditions.
Four steps. Then you're working.
No config files. No Docker. No SSH. Sign up, paste a key, pick a channel, go.
Sign up
Email or Google, ten seconds.
Connect a model
Paste an Anthropic, OpenAI, Google, or OpenRouter key. BYOK.
Add brand context
A few lines about your business — what you do, voice, audience.
Connect a channel
Telegram in 2 minutes. Or chat right in the dashboard.
A calm cockpit for your agent fleet.
Chat. Voice. Goals. Mission Control. Approvals. Skills. Channels. Brand context. Memory. Schedule. Files. Everything one click away — designed for humans, not engineers.
Keeps summaries, follow-ups, and business memory organized so nothing falls through the cracks.
Captured contact details, asked follow-ups, and flagged the highest-intent buyer.
Summarized updates, open loops, and next actions from files and chat context.
Drafted a response, but waited for approval before sending externally.
Simple pricing. BYOK.
You pay Zexo for the platform. You pay your model provider directly for tokens — capped by your spending limit so there are no surprises.
Your first agent. Solo operator, all the guardrails.
- 1 private workspace
- Telegram + Discord (webhook) + Voice + Web chat
- 13 default skills + MCP server connections
- Tiered memory + brand context
- Mission Control + replay viewer
- Spending caps + approval queue + kill switch
Real operator workflows. All channels, scheduled work, sub-agents.
- Everything in Starter plus:
- All channels including Discord bot (/zexo command)
- Scheduled tasks (cron-pinned proactive runs)
- Sub-agent orchestration (delegate focused tasks)
- Workspace files (Supabase storage)
- Up to 3 seats
Teams & agencies. Multiple parallel goals, custom integrations.
- Everything in Pro plus:
- Up to 10 seats
- Custom MCP server connections (your own internal tools)
- Goals kanban (multi-project tracking)
- Higher rate limits
- Priority support
Questions, before you start.
The things people ask. If something else is on your mind, reach out — we'd rather you know than guess.
What does my AI agent actually do?
Whatever you set it up to do. Common jobs: triage email, send follow-ups, post to Slack/Discord, run morning briefings, draft content, qualify leads, summarize calls. You describe the work, it does the work, you see every step in Mission Control.
Do I need technical skills?
No. No Docker, no terminal, no scripts. Sign up, paste an API key, connect a channel. The whole product is browser-based.
How is this different from OpenClaw or Hermes-Agent?
OpenClaw and Hermes are powerful CLI tools that need 15+ hours to set up properly and have no built-in spending caps (users have reported $5k overnight runs). Zexo is a hosted SaaS with budget guardrails, an approval queue, and live Mission Control built in. Same agentic power, none of the setup tax or runaway-cost risk.
Which model providers do you support?
Anthropic (Claude), OpenAI (GPT), Google (Gemini), and OpenRouter. Bring your own key for any of them. Pick a preferred provider in Settings; if it's down or you switch, the agent gracefully falls back.
Where does my data live?
Each paying customer gets an isolated workspace. Your conversations, files, and memory stay yours. No shared pools, no training on your data, no exposure to other customers. API keys are encrypted at rest with AES-256-GCM.
Can I plug in my own tools?
Yes — through Model Context Protocol (MCP). Any MCP-compatible server (yours or someone else's) shows up as new skills the agent can call. You point at a URL, optionally add auth, and we discover the tools automatically.
What if the AI does something wrong?
Approval-required skills (like sending email) never run automatically — they queue in /dashboard/approvals and wait for you. Spending caps stop runaway runs. Kill switch pauses everything in one click. Mission Control replays exactly what happened so you can see why.
Can I cancel?
Anytime, from your dashboard. No contracts. Your workspace stays read-only for 30 days in case you change your mind, then it's fully deleted.
Your agent is ten minutes away.
Paste a model key. Connect a channel. Start delegating the work that's been piling up.