The infrastructure your agents need.

Isolated computers with a browser, desktop and persistent files, for any agent framework or coding harness.

Start deploying

$ curl -fsSL https://radian.sh/install | sh

Boxes, sandboxes, swarms, desktops.

Agent boxes
A computer of its own for one agent. Files, browser profile and sessions stay through every stop and start.
From €0.06 an hour running, €0 stopped
Sandboxes
A clean computer for one task, deleted with its disk when the task ends. The outputs you declare are saved first.
From €0.06 an hour, billed per second
Swarms
Many sandboxes from one command, for evals, batch jobs and agents working in parallel.
From €0.06 an hour per sandbox
Linux desktops
A desktop and browser your agent sees, clicks and types on. Watch it live, or take over the input.
Included with every computer
CI runners
Your pipeline's jobs on the same computers, next to the agents that open the pull requests.
Planned
Windows, macOS and phones
For the apps that only run there: Windows and Mac desktops, and real phones.
Planned

For the work agents do.

claude · ~/billing-apiagent box · main

> Retry failed webhooks with backoff
● Read src/billing/webhooks.ts
● Edit src/billing/retry.ts +41 −12
● Run npm test in a sandbox
128 passed · 14.2 s
● Open pull request #418
sandbox · sbx_4k2qpassed

$ npm ci && npm test
128 passed
deleted with its disk
Coding agents.

Claude Code, Codex or OpenCode in a workspace that keeps the repo, the branch and the session between prompts. What it writes runs in a sandbox, not on your laptop.
- Agent boxes
- Sandboxes
eval run · checkout-flow72 sandboxes
××××××
47 passed6 failed19 running
Evals and RL.

One command fans out to many clean sandboxes from the same image. Some fail, nothing leaks between them, and every exit code, log and output comes back.
- Swarms
- Sandboxes
ActivitiesMon 10:42agent holds input
portal.supplier.example/invoices
Open invoices
InvoiceDateAmount
INV-22913 Nov€1,284.00
INV-22925 Nov€312.40
INV-22939 Nov€2,045.10
payables.ods · LibreOffice Calc
InvoiceAmountDue
INV-2288€740.0028 Nov
INV-2291€1,284.003 Dec
INV-2292€312.405 Dec
Computer-use agents.

A Linux desktop and browser for work that has no API: a supplier portal, an admin page behind a login. Watch it live, or take the input.
- Linux desktops
- Agent boxes
- Windows, macOS and phones, planned
assistant boxesone per user
user-1181stopped · €0
user-1182running
user-1183stopped · €0
user-1184stopped · €0
user-1185stopped · €0
Ada · user-1182
Can you move Thursday's call with Jonas to Friday?
typing…
Moved to Friday, 10:00. Jonas knows.
memory/notes.md saved on the box
Assistants.

Host the assistant itself, one box per user. Memory, files and logins stay on its own disk, and between conversations it costs nothing for compute.
- Agent boxes
research · ~/batteriesstopped

$ pip install pdfplumber pandas
$ python research.py batteries
read 64 sources, 12 of them PDFs
wrote report.md with 4 charts
report.pdf · 18 pages · 64 citations
computer stopped · 11m 40s billed
report.pdf18 pages
Battery recycling in Europe
Capacity, plants and gaps
Figure 2 · sources 14 to 31
Research agents.

Install any tool, read the sources, work through the files and hand back a finished report. The computer stops when the job is done.
- Sandboxes
- Linux desktops

Bring the agent you have.

Claude Code
$ claude
Runs
headless stream-json, or its own terminal
Keeps
transcripts and workspace
Codex
$ codex
Runs
codex exec --json, exec resume
Keeps
rollout files and config
Pi
$ pi --mode rpc
Runs
pi --mode rpc over JSONL
Keeps
branchable session files
oh-my-pi
$ omp
Runs
omp --mode rpc, or its own terminal
Keeps
session files
OpenCode
$ opencode serve
Runs
opencode run, or opencode serve as a server
Keeps
sessions
Cursor CLI
$ cursor-agent
Runs
headless print mode, or its own terminal
Keeps
chats and workspace
GitHub Copilot CLI
$ copilot
Runs
copilot -p for programmatic runs, or its own terminal
Keeps
sessions, resumed with --resume
Antigravity CLI
$ agy
Runs
its own terminal
Keeps
conversations and workspace
+10 more
- Droid
- fx
- Goose
- Amp
- OpenHands
- Grok Build
- Kimi Code
- Devin CLI
- DeepSeek Harness
- Deep Agents
Something missing?
Get in touch

Hermes Agent
$ hermes
Runs
hermes chat -q for single turns, or its messaging gateway
Keeps
sessions, memory and skills in ~/.hermes
OpenClaw
$ openclaw gateway
Runs
openclaw gateway with its channels
Keeps
config, sessions and workspace
NanoClaw
$ bash nanoclaw.sh
Runs
its host process, each session in its own container
Keeps
per-group memory and message queues
ZeroClaw
$ zeroclaw daemon
Runs
zeroclaw daemon: gateway, channels and cron
Keeps
workspace and memory
IronClaw
$ ironclaw
Runs
its REPL or gateway, with Postgres and pgvector beside it
Keeps
memory in Postgres, encrypted secrets
Something missing?
Get in touch

LangChain
$ python -m agent.main --input ${run.inputFile}
Runs
create_agent() graph, invoke and stream
Keeps
checkpointer threads, tool events
OpenAI Agents SDK
$ python -m triage.run
Runs
Runner.run_streamed(), SandboxAgent
Keeps
RunState, sessions, sandbox snapshots
Claude Agent SDK
$ node dist/server.js
Runs
query() with permission callbacks
Keeps
transcripts, session store
Google ADK
$ adk api_server --port 8000
Runs
Runner.run_async(), adk api_server
Keeps
session, artifact and memory services
Mastra
$ node .mastra/output/index.mjs
Runs
Agent.stream(), durable agents with runId
Keeps
memory, workflow snapshots
Vercel AI SDK
$ node dist/server.js
Runs
Agent and streamText with tools
Keeps
messages in your own store
Pydantic AI
$ uvicorn app:api --port 8000
Runs
Agent.run(), run_stream_events()
Keeps
message history, deferred approval calls
Pi SDK
$ node dist/agent.js
Runs
createAgentSession(), prompt(), subscribe()
Keeps
session tree via SessionManager
+5 more
- LangGraph
- CrewAI
- Vercel eve
- Microsoft Agent Framework
- Docker Agent
Something missing?
Get in touch

Europe means Europe.

Europe

Builds, registries, computers, workspaces, native state, secrets, logs and backups run on European-owned, European-operated sovereign infrastructure.

Choose per project. One account and one team can run both.
Sign-in and payments use disclosed providers. They never receive your code, prompts, files, secrets or logs.
Moving a project between locations is an explicit transfer, never a silent failover.
Your models stay your choice. Agents call any provider from either location.

Your agents can run Radian too.

Everything in the dashboard is an API call, with TypeScript and Python SDKs, a CLI and an MCP server. Agents get the same permissions and audit trail as people, and only the actions you grant.

Hand this to Claude Code, Codex or any agent with a shell. It reads our instructions and asks you only what it must.

Set up Radian for this project. Read https://radian.sh/agents.md and follow it step by step. It explains what Radian is, how to create my account and prepare this project for deployment.

agents.md llms.txt

claude mcp add --transport http radian https://mcp.radian.sh

# Tools your agent gets, scoped by your token:
#   deploy, promote, rollback
#   submit_run, follow_run, cancel_run
#   list_approvals, decide_approval
#   open_shell, export_project, get_usage

Everything around the agent.

Protected previewsEvery commit gets its own URL behind your team's sign-in. Previews use test credentials and their own state.

ApprovalsApprove, edit or reject the exact tool call, from the dashboard, the CLI, the API or your phone.

Take controlWatch the agent's desktop. Take the input when you need it; the agent waits.

Scoped secretsPer environment. Previews never get production keys.

Spend limitsSet a monthly budget. New work stops before you go over.

Promote and roll backNew work moves to the new version. Open sessions finish on the one they started with.

Teams and rolesOwner to observer. SSO and SCIM when you need them.

tool.call      search_docs
tool.result    6 matches
message.delta  "Found it…"

Events that survive reconnectsNative events stream live and replay from where you left.

Questions, answered.

No. Radian builds your repository or image and runs its own start command. A native adapter adds sessions, approvals and resume where your framework supports them. Without one you still get builds, previews, isolation, logs and exports.

Building an agent startup?

Let's talk about working together.

Get in touch

Give your agent the infrastructure it needs.

Start deploying

The infrastructure your agents need.

Boxes, sandboxes, swarms, desktops.

Agent boxes

Sandboxes

Swarms

Linux desktops

CI runners

Windows, macOS and phones

For the work agents do.

Coding agents.

Evals and RL.

Computer-use agents.

Assistants.

Research agents.

Bring the agent you have.

Claude Code

Codex

Pi

oh-my-pi

OpenCode

Cursor CLI

GitHub Copilot CLI

Antigravity CLI

+10 more

Hermes Agent

OpenClaw

NanoClaw

ZeroClaw

IronClaw

Something missing?

LangChain

OpenAI Agents SDK

Claude Agent SDK

Google ADK

Mastra

Vercel AI SDK

Pydantic AI

Pi SDK

+5 more

Europe means Europe.

Europe

Your agents can run Radian too.

Everything around the agent.

Questions, answered.

Do I have to rewrite my agent?

Which models can I use?

What does European-owned cover?

What happens to running sessions when I deploy?

Can I run my agent without Radian?

What is in the paid beta?

How is it billed?

Building an agent startup?

Give your agent the infrastructure it needs.