CelestoAI

Computer for AI Agents

[CelestoAI @CelestoAI]

Huge thanks to @khushbooverma_ for speaking at our meetup 🧡

[Khushboo Verma @khushbooverma_]

delivered a talk on building runtime boundaries with docker sandboxes in london last evening

give the agent total autonomy, but explicitly control the blast radius

thanks for having me at the enterprise ai agents meetup, @CelestoAI 🚀

[Khushboo Verma @khushbooverma_]

delivered a talk on building runtime boundaries with docker sandboxes in london last evening

give the agent total autonomy, but explicitly control the blast radius

thanks for having me at the enterprise ai agents meetup, @CelestoAI 🚀

[Aniket @aniketmaurya]

Finance is one of the best places to deploy AI agents.

Not because it’s easy.

Because the workflows are painful, repetitive, high-value, and trapped inside legacy systems.

@therealzling is speaking at our London AI agents meetup on the real challenges of deploying agents in regulated industries.

Focus: finance.
Also touching legal and healthcare.

Zhi has worked across fintech, qualified as a barrister, spent time in the NHS, and previously built a healthtech startup.

She’ll cover use cases, bottlenecks, and how teams actually get agents into production.

Join us: https://t.co/7qmkk9yACH

[Aniket @aniketmaurya]

.@khushbooverma_ from @Cloudflare is speaking at our London AI agents meetup.

The future of agentic development is not just better models.

It’s better sandboxes.

Once agents can run commands, install deps, use Docker, call tools, and make network requests, they become a runtime security problem.

Khushboo is also a @Docker Captain and technical creator followed by 160K+ developers.

She’ll demo how to run agents in bounded environments without handing them the keys to the host.

🎟️ Come through: https://t.co/7qmkk9yACH

[Aniket @aniketmaurya]

London is one of the financial hubs of the world.

That makes it the perfect ground for AI agents in regulated industries.

Come learn about computer-use agents and their applications in finance, back-office operations, and legacy software.

https://t.co/ocdt6BI67f

[CelestoAI @CelestoAI]

@venkat_systems @jonobelotti_IO @aniketmaurya ✴️

[CelestoAI @CelestoAI]

We’re launching Celesto for computer-use agents.

Give your agent a computer.
Let it open software, click around, fill forms, and complete real workflows.

This is how AI moves from chat → work.

Demo below 👇

[Aniket @aniketmaurya]

RPA was built for deterministic workflows.

Celesto is built for agents that can think, adapt, and use computers.

We’re launching computer-use agents that can operate software like a human:
Click, type, read, recover, and finish the job.

Watch Celesto automate anything 👇

[Aniket @aniketmaurya]

RPA was built for deterministic workflows.

Celesto is built for agents that can think, adapt, and use computers.

We’re launching computer-use agents that can operate software like a human:
Click, type, read, recover, and finish the job.

Watch Celesto automate anything 👇

[CelestoAI @CelestoAI]

@aniketmaurya 🚀

[Aniket @aniketmaurya]

Just shipped a prebuilt Hermes agent sandbox.

Boots in seconds. Runs locally. 🚀✴️

Just run: smolvm hermes start

[Aniket @aniketmaurya]

every coding agent gives you two bad options:

→ approve every command (death by prompt)
→ skip permissions and pray

so I built smolvm. local microvm sandbox for Claude Code, Codex, and Pi.

one command. full permissions. zero host risk.

free + open source 👇

https://t.co/kdNbIiw8OC

[CelestoAI @CelestoAI]

You can start Claude or Codex with a mounted workspace in a single line.

smolvm codex start --mount ./folder-path

[Aniket @aniketmaurya]

Launching SmolVM -- run claude and codex in sandboxed environment with full permissions worry free.

It sucks to “press enter and accept changes” every few seconds while using coding agents. We made it easy to isolate the agent coding environment from the host (laptops).

With a single command you get a claude/codex pre-installed sandbox ready with git credential to make you build a billion dollar business without making any mistake 😉

To give it a try, head to the open-source repo and install SmolVM then run: "smolvm codex start" or "smolvm claude start --mount ./my-project" for a quick experiment.

[CelestoAI @CelestoAI]

https://t.co/NmH8X9RqT4

[CelestoAI @CelestoAI]

Managed Claude Agents

[Aljosa Asanovic @aljosa]

Celesto AI

Managed agent platform on top of their open-source SmolVM engine. Agent framework with MCP support, agent-to-agent communication, Tool Search API to reduce context bloat, and delegated OAuth for end-user resources.

https://t.co/YqTySG1LsA

Founders: @aniketmaurya

[CelestoAI @CelestoAI]

Managed Claude Agents but for any LLM provider

[Aljosa Asanovic @aljosa]

Celesto AI

Managed agent platform on top of their open-source SmolVM engine. Agent framework with MCP support, agent-to-agent communication, Tool Search API to reduce context bloat, and delegated OAuth for end-user resources.

https://t.co/YqTySG1LsA

Founders: @aniketmaurya

[Aljosa Asanovic @aljosa]

Celesto AI

Managed agent platform on top of their open-source SmolVM engine. Agent framework with MCP support, agent-to-agent communication, Tool Search API to reduce context bloat, and delegated OAuth for end-user resources.

https://t.co/YqTySG1LsA

Founders: @aniketmaurya

[Aniket @aniketmaurya]

Not just Claude but with any agent with @CelestoAI

- Our sandboxes are open-source and based on firecracker -- spin up VMs in milliseconds to use as tool or run the agent inside the VM.

- Clone the whole environment in seconds.

- Spin up thousands of machines super fast ⚡️

[CelestoAI @CelestoAI]

We just made smolvm ssh instant for running VMs.

Before: ~15s
After: milliseconds

What went wrong?

The CLI was probing SSH readiness every time:
polling ports, trying handshakes, and burning timeout budget before connecting.

That makes sense when a VM is still booting.
It makes no sense when the VM has already been running for minutes.

So we fixed it:
for running VMs, we skip the probe, build the SSH command from known host/port/key info, and connect directly.

Sometimes the best performance fix is just:
stop doing unnecessary work.

[Aniket @aniketmaurya]

A sandbox that can reach any domain on the internet is not much of a sandbox.

We just added domain allowlists to SmolVM.

[CelestoAI @CelestoAI]

One step, for an instant secure agent infra -- "pip install smolvm".

[Aniket @aniketmaurya]

Stop using Docker as a sandbox!!
Agents run code, and too many setups still execute untrusted workloads with weak isolation.

That’s why we built SmolVM — an instant, secure computer for AI agents.

Built for AI agents:
🔒 hardware level isolation (built for untrusted workloads)
🛜 Network isolation: URL allowlisting and toggle internet access
📂 Elastic file storage system
⚡️ fast boot + low overhead

and much more... Check the link to the GitHub repo.