Agents Made Simple, Secure.
The Platform. Built for Humans.
Create, Discover
Build with natural language on any model. Rapidly prototype, version, and iterate. Discover Experts in the Gallery powered by MCP and Agent Skills.
Run, Deploy
Open-source sandboxed runtime. Secure isolation with full autonomy. Bring your own providers and storage. Deploy instantly via Serverless Sandbox API.
An Agent-First Platform to Build Better Agents.
The Problem Autonomy without focus is chaos.
Autonomy without focus is chaos.
The 'do everything' agents proved it.
Early agents promised to handle any task. They ended up looping forever, hallucinating actions, causing irreversible damage. The survivors became chat interfaces—safe, but generic. The real answer isn't giving up on autonomy. It's giving agents focus.
The Essence Model. Prompt. Tools.
Model. Prompt. Tools.
That's all an agent is.
Frameworks add orchestration, memory, application logic. None of it makes your agent smarter, more reliable, or more capable. An agent's quality comes from the model you choose, the prompt you write, and the tools you give it. Everything else is noise—it helps you ship, but it doesn't help your agent perform.
The Answer Expert = agent, focused.
Expert = agent, focused.
Autonomy through focus, not despite it.
Claude Code didn't give up on autonomy—it earned it through focus. One purpose (coding), one interface (CLI), one goal at a time. An Expert works the same way. Focus isn't a limitation. It's how agents become trustworthy.
The Stack The Expert Stack.
The Expert Stack.
Everything an Expert needs. Nothing it doesn't.
Define your Expert in plain text—instruction, skills, purpose. The open-source runtime executes it safely. The gallery makes it discoverable and reusable. You focus on the essence. The stack handles the rest.
Create. Write agents, not applications.
No framework to learn. No boilerplate to write. Just instructions, skills, and permissions.
Agent-first
Write instructions. Add skills. Assign permissions. The environment for your masterpiece is ready.
Zero friction for users
No installation guides needed. Users run your agent safely, instantly. Just share the definition and best practices.
Feedback-driven growth
Lower barriers mean more users. More users mean more feedback. Real human feedback is the fastest path to better agents.
[experts."my-agent"]
instruction = "You are a helpful assistant..."
skills = ["github", "slack"]
permissions = ["read:repo", "send:message"] Exhibit. With Confidence.
Every agent in the Gallery is audited. Every execution is sandboxed.
Sandboxed
Every agent runs isolated. User system stays safe.
Audited
LLM-powered audit checks every public agent.
Transparent
Know exactly what an agent can access.
Run. Open source isn't optional.
Nothing hidden. Nothing proprietary. Nothing you can't verify.
$ npx perstack start \
--runtime docker \
--workspace path/to/workspace \
--model claude-4.5-opus \
bug-finder \
"Find 3 bugs in packages/runtime" // visible Nothing hidden
Agents fail. When they do, you need to know exactly what happened. Every action logged. Every state visible. No guessing.
// isolated Sandboxed. Obviously.
Agents delete databases. They hallucinate commands. Isolation isn't optional—it's the baseline.
// standard Built on what you already trust
Docker. Node.js. Standard tools. No proprietary runtime to learn. No vendor lock-in to escape.
Improve. Together, through collaboration.
Better agents don't come from better models. They come from the right people working together.
Domain experts define
Write what the agent should do—in plain language. The closer you are to the domain, the better the agent.
Users refine
Use it. Report what works, what doesn't. No automated evaluation can replace real feedback.
Engineers integrate
Embed agents into applications. Standard tools, UNIX principles—composable, predictable, auditable.
Three roles. Three concerns. One foundation.
"I'm not a developer. Can I still build agents?"
Plain language. No framework. Focus on what you know.
"How do I know this won't break things?"
Sandboxed. Explicit permissions. Full audit trail.
"Another proprietary runtime to learn?"
Docker. Node.js. Open source. Nothing proprietary.