The cloud where the software your agent builds lands.
Your agent deploys the tool it just built in one call. Your team opens it like a Google Doc. It runs sandboxed, and it is yours to take with you.
Agents can write a bespoke internal tool in seconds. Deploying it, securing it, and getting it to your team is still the hard part. Perch is that part.
Three hard problems, gone.
Runs untrusted code, safely
Every tool executes in a locked-down sandbox with no filesystem, network, or host access. Its only capability is its own isolated storage, handed in by the platform. It shipped with an adversarial security review that found a real escape, now fixed and tested against.
Auth and permissions, built in
Tools inherit your org identity. The platform authorizes every request before a line of tool code runs. Share by person, group, org, or public, and it plugs into your SSO.
Shared like a doc, owned like a file
Send a link the way you send a Google Doc. And export the exact source whenever you want. Nothing is trapped inside Perch.
One call, from the agent to a running tool.
No dashboard, no deploy vocabulary, no cloud console. The agent does it while it builds.
Agent ships it
Over MCP or the CLI, the agent deploys the tool it wrote and gets a URL back.
Sandboxed
It runs isolated, with its own storage. Cold by default, so an idle tool costs nothing.
Like a doc
Private by default. Share with a person, a group, your whole org, or the public.
Yours to take
Export the exact source anytime. It runs anywhere. No lock-in.
The agent is the user of this cloud.
Perch is a small API over MCP and a CLI. Point your coding agent at it and it deploys and shares the software it just wrote, without leaving its environment.
Six MCP tools
Deploy, list, share, read logs, read source, sign in. Wired in one line.
Redeploy in place
The agent iterates on the same URL and reads its own source back to revise it.
Human-friendly on the other side
Your teammates get a clean, permissioned page. They never see the plumbing.
It runs untrusted code, and says so plainly.
The sandbox is defense in depth: a vm with a deny-all import linker and no host-realm references, inside a process locked down by the OS, with storage kept out of reach. It shipped with a written threat model, not a promise.
On the hosted service, we own the hardest part: multi-tenant isolation, upgrades, and keeping the sandbox ahead of the threat model, so your team never has to think about it. The source is available to read and evaluate, and we are direct about the boundary rather than overclaiming it.
Let us run it for you.
Perch is source-available to read and evaluate. For your team, the hosted service is the way to use it: managed isolation, SSO, and support, with none of the hosting to maintain.
Perch for your team
We run it. You share tools. Managed sandbox, SSO, upgrades, and support.
- Managed multi-tenant isolation, kept hardened
- SSO with your identity provider
- MCP + CLI for every agent on your team
- Export any tool to source, anytime
Early access. We onboard teams one at a time.
Self-evaluate
Read the code and run it locally for personal, noncommercial use.
- Full source, clone and run in one command
- The complete test suite and threat model
- PolyForm Noncommercial license
- Commercial use goes through the hosted plan