Inkbox gives AI agents their own cloud computers and phone numbers
Inkbox just launched hosted agents alongside a $5.3M seed round, giving AI bots dedicated cloud computers, phone numbers, and inboxes. The agents sleep when idle to cut compute bills and wake up automatically whenever a lead or customer replies. It turns flaky one-off scripts into persistent, reachable digital employees.
If you have ever tried deploying an AI agent for real work, you know the awkward reality of async tasks. You prompt an agent to shoot out cold emails, it fires them off via an API, and the execution terminates. When a prospect writes back four hours later, your agent is already dead, completely unaware that a conversation just started.
To keep it reachable, developers have traditionally faced two equally painful options: keep expensive cloud instances idling around the clock, or spend days wiring up webhooks, database states, and wake-up logic across fragmented tools like Twilio, Gmail, and ngrok.
Y Combinator alum Inkbox is stepping up to eliminate that plumbing. Alongside a fresh $5.3M seed round led by Peak XV, the startup has launched hosted agents: AI bots that come packaged with their own persistent cloud computer and real-world communication handles.
A persistent computer with a real phone number
Inkbox was founded by MIT CS PhD Ray alongside Alex and Dima, who previously built machine learning pipelines at Harvard Medical School. Backers in the round include Vercel CEO Guillermo Rauch, Databricks Chief AI Scientist Jonathan Frankle, Linear COO Cristina Cordova, and Accel partner Dan Levine.
The founders ran into this exact headache while building web-automation bots. Every time their agents tried logging into third-party portals, they got stopped by confirmation emails and two-factor codes. Over the four months since their YC debut, Inkbox has already routed millions of messages for AI agents looking for persistent identities.
Now, they are taking over the compute layer as well. Hosted agents don't just get an email address; they get a full Linux computer tied to that identity. You can drop in an agent harness like Hermes, Codex, Claude Code, or OpenClaw, open a terminal, install dependencies, and save files to disk.
Crucially, the agent gets dedicated inboxes across email, phone, iMessage, WhatsApp, Slack, and even agent-to-agent channels. There is even a Grok Bot plugin ready to go.
Sleep until spoken to
The real secret sauce here is the sleep-and-wake architecture.
Running tasks 24/7 gets expensive quickly. On Inkbox, when an agent finishes sending out messages or running a scheduled cron job, it goes to sleep. You do not pay for idle compute time.
Yet the moment an email reply or an inbound text hits its inbox, the platform wakes the machine up. Because its disk state, file system, and conversation threads remain untouched, the agent resumes right where it left off, piecing together the full context without developer intervention.
Developers discussing the release on daily.dev have pointed out obvious benefits, particularly for long-running customer support or sales workflows. But the community also flagged immediate trust and security challenges: what happens when an unknown sender attempts prompt injection over email, or spam messages cause constant, costly false wake-ups?
Why it matters
We think Inkbox is tackling one of the most glaring infrastructure bottlenecks in practical agent deployment.
While the industry has spent the last year racing on frontier model intelligence and orchestration frameworks, almost nobody solved the mundane reality of running these bots in the wild. The moment an agent has to interact with humans across multiple days, you end up stitching together legacy APIs meant for human workers. Inkbox turns that messy plumbing into a clean, integrated utility.
For builders and small teams, this dramatically lowers the barrier to deploying autonomous SDRs or asynchronous support reps. Instead of maintaining sprawling webhook servers, you hand your prompt a computer and a phone number.
That said, giving an autonomous agent a live shell and open communication lines brings serious trust-and-safety liabilities. Inkbox provides scoped permissions and contact rules, but treating incoming emails as untrusted inputs while preventing hallucinated out-of-bounds actions is still an unsolved problem.
Even so, the direction is unmistakable. Agents are no longer just ephemeral chat widgets; they are becoming reachable entities with state, disks, and contact cards. Inkbox just gave them an office desk.