Self-hosted environments for the agents you already use, without the hosting.
No credit card required. 7-day free trial.
Supported providers
- Claude Code
- Claude Managed Agents
- OpenAI
- Devin
Amp
The self-hosted option, without the ops work.
Claude, OpenAI, Devin, and Amp can run sessions on compute you choose. Doing that yourself means deploying a worker, keeping machines warm, and patching a runner image. Connect your provider to Depot instead, and every session gets a fresh sandbox.
self-hosted environment / ops
Deploy the provider’s worker
Handled
Keep machines warm for new sessions
Handled
Patch the runner image
Handled
Scale for parallel sessions
Handled
Clean up after every session
Handled
- Workers to run
- 0Workers to run
- Sandbox per session
- 1Sandbox per session
- Agent providers
- 5Agent providers
Any agent, no workers
Connect with credentials from your provider. Depot picks up new sessions, boots a sandbox for each one, starts the provider’s runner, and cleans up when it’s done.
Agents build as fast as your builds
Sandboxes run on the same Depot Metal machines as Depot CI, with Depot Cache and your Tailscale tailnet a setting away.
Controls your security team can review
Host-bound secrets never enter the sandbox. Runs reach only the repositories you pick. Every session runs in its own short-lived microVM.
Bring the agent you already use.
Your provider keeps the conversation, the model, and your prompts. Depot supplies the machine. Connect once, then start sessions the way you already do.
Claude Code sessions, on Depot.
Start cloud sessions from Claude on the web, desktop, or mobile, and route them to a self-hosted environment connected to Depot. The agent loop and its tools both run in the sandbox.
Requires a Claude Team or Enterprise plan with self-hosted environments enabled.
Read the Claude Code guideclaude code / setup
01
Create a self-hosted environment in Claude Code and copy its key.
02
Paste the key into Depot under Agent providers.
03
Start a session in Claude Code with that environment.
What happens when you start a session?
Depot claims the session from your provider, boots a sandbox, and starts the provider’s runner. When you return to the conversation, the next sandbox starts with the agent’s files in place.
Claude Code
Claude Managed Agents
OpenAI
Devin
Amp
One microVM per session
Depot sandbox
A microVM on Depot Metal with Depot Cache, your tailnet, and your run settings.
Start sandbox
Run setup script
Run agent
Workspace snapshot
Files restored when you come back
Managed Runs
Status, steps, logs, and outputs
01
Your provider queues a session.
02
Depot boots a sandbox with your run settings.
03
The agent works while Depot snapshots its workspace.
04
Come back later and its files are still there.
See every run, whoever started it.
Sessions from every connected provider land in one list, so the whole team can see what’s running and why a run failed.
Filter runs by provider and status
Step durations and logs, including your setup script’s output
Download files the agent saved to /workspace/outputs
Stop a run that’s still working
Managed Runs
All providers
Claude Code
OpenAI
Devin
Amp
Claude Code
3m 12s
Running
OpenAI
6m 48s
Succeeded
Devin
12m 05s
Succeeded
Claude Managed Agents
41s
Failed
The agent exited with an error before the run finished.
Amp
2m 30s
Cancelled
OpenAI run
6m 48s
Steps
Start sandbox
1s
Run setup script
14s
Run agent
6m 33s
Outputs
summary.md
4.2 KB
test-results.xml
18 KB
Open in OpenAI
Every session starts with your settings.
Configure a connection once. Its run settings apply to every session on it, whichever provider started it.
Run secrets
Host-bound secrets go only to the hosts you list, so they never enter the sandbox.
GitHub access
Pick which repositories runs can read or write. Runs clone and push through Depot’s Code Access app.
Private services
Sandboxes join your Tailscale tailnet, so agents can reach internal databases and APIs.
Setup script
Install tools before the agent starts. The script’s output shows up as its own step on the run.
MCP servers
Give every run on a connection your organization’s MCP servers, with nothing to set up per session.
Sandbox size
Choose 2, 4, 8, or 16 vCPUs per connection, with 4 GB of memory each. The default is 4 vCPUs.
$0.36 for a 30-minute agent session.
That’s the default 4 vCPU, 16 GB sandbox, billed per second at $0.00005 per vCPU-second. The same rate as Depot CI and Depot Sandboxes.
Warm between replies
Claude Code, Claude Managed Agents, and OpenAI sandboxes stay up for five to ten minutes after the agent goes quiet. Reply in that window and the agent picks up on the same machine.
Included minutes cover agents too
Managed Agents draws from the same included sandbox minutes as Depot CI. The Startup plan includes 20,000, measured at 2 vCPUs.
No charge for tokens
Depot bills for the sandbox, including that idle window, at the same per-second rate. Model usage and subscriptions stay on your provider account.
Good to know.
Resources
Learn more
about Depot Managed Agents.