"The final barrier between enterprise security and AI-native development just collapsed. By moving the agent execution inside the perimeter, Cursor isn't just a tool anymore—it's part of your private infrastructure."
For years, "Vibe Coding"—the practice of shipping production-grade software primarily through high-level intent and AI orchestration—hit a hard ceiling: The Security Review. You could build a world-class MVP in a weekend, but the moment you moved to a high-compliance environment (Fintech, Healthtech, or internal Big Tech monorepos), the "cloud" part of "AI Cloud Agents" became a non-starter.
That changed on March 25, 2026. Cursor’s announcement of Self-Hosted Cloud Agents allows you to run the same "God Mode" capabilities of their hosted agents entirely within your own network. Your code, build outputs, and—crucially—your secrets never leave your infrastructure.
This guide is for the builder who wants to leverage this "Sovereign AI" stack to ship faster while keeping the CISO happy.
Why this matters for builders
Self-Hosted Cloud Agents let you orchestrate complex coding tasks across your codebase using your own private infrastructure.
Until now, using an AI agent meant sending context (and sometimes your entire repo) to a third-party sandbox. This update unlocks three massive shifts:
- Zero-Egress Compliance: You can now use AI on sensitive PII, proprietary algorithms, or air-gapped internal tools.
- Local Secret Management: The agent can interact with your internal DBs, private APIs, and staging environments because it lives inside the VPC.
- End-to-End Automation: You can trigger these agents from Slack, Linear, or GitHub to perform tasks like "Fix this bug and deploy to our internal staging," with the agent executing the build on your own hardware.
When to use it
Not every project needs a self-hosted agent. Use this setup when:
- You work in a regulated industry: HIPAA, SOC2, or GDPR requirements prevent code from leaving specific regions or networks.
- You have a massive monorepo: Sending gigabytes of context to the cloud is slow and expensive; local agents have "home field advantage."
- You need "Hardware-in-the-loop": If your build requires specific GPU clusters or local dev-kits that aren't available in standard cloud sandboxes.
- You want a "Sovereign Dev Stack": You prefer having total control over the agent's uptime, resources, and plugins.
The Full Process: From Zero to Self-Hosted Agent
To ship something real with this, we’re going to walk through the "Vibe Coding Workflow": Scoping, Prompting, Scaffolding, and Shipping.
1. Scoping: Define the Agent's Territory
Before turning it on, decide what the agent is actually allowed to do. In the self-hosted world, the agent has the "keys to the house."
- The Goal: Create an agent that monitors your internal Linear board for "Bug" tags, reproduces them in your local staging environment, and submits a PR.
- The Spec: The agent needs access to your internal GitHub/GitLab, your local Docker registry, and your Linear API keys (stored in your internal vault).
2. Scaffold: Setup the Infrastructure
According to the March 2026 announcement, you enable this via the Cursor Dashboard.
- Enable: Go to your Cursor Dashboard and toggle "Self-Hosted Cloud Agents."
- Provision: You will need to provide a compute target. While Cursor handles the orchestration, the agent executes in isolated VMs or containers within your network.
- Connect: Use the Cursor API or CLI to link your internal runner to the Cursor control plane. (Note: Check the official docs for the exact
curlordocker composecommand for your specific OS).
3. Prompt: The Sovereign Spec
When "Vibe Coding" with a self-hosted agent, your prompts change. You no longer need to worry about "Is this secret safe?" Instead, you guide the agent through your local network topology.
Copy-paste this starter template for a Sovereign Agent Prompt:
### AGENT ROLE
You are a Senior Infrastructure Engineer operating within our private VPC.
### CONTEXT
- Codebase: Located at /src/internal-app
- Build Tools: We use internal Jenkins and a local Artifactory.
- Security: All secrets are in the local environment variables. Do NOT attempt to reach out to external public APIs unless explicitly listed.
### TASK
1. Monitor the Linear webhook for ticket #BUG-123.
2. Spin up a local development container using `make dev-shell`.
3. Identify the failing test in `tests/integration/`.
4. Fix the code, verify the build passes locally, and push a branch to our internal GitLab.
### CONSTRAINTS
- Use only internal documentation found in /docs/internal-wiki.
- Do not log any PII found in the database.
4. Implement: Execution and Interaction
Trigger the agent. You can kick this off from the Cursor editor, but for true Vibe Coding, you’ll likely trigger it via Slack or GitHub Actions.
- Vibe Check: Watch the agent's progress in the Cursor Web App. Because it's self-hosted, you should see logs appearing in your own observability stack (Datadog, Grafana, etc.) alongside the Cursor UI.
- Tool Usage: Since self-hosted agents support plugins, you can give the agent custom tools to interact with your proprietary internal APIs.
5. Validate: The Trust-but-Verify Loop
Since the agent is in your network, validation is easier:
- Audit Logs: Check your network traffic. Ensure the agent isn't attempting to call home with code snippets.
- Build Integrity: Ensure the agent's "build outputs" match your standard CI/CD artifacts.
- Isolated VMs: Confirm that the agent’s VM is wiped after the task is complete—a key feature of the Cursor-hosted version that is preserved in the self-hosted release.
Pitfalls and Guardrails
What if the agent gets stuck in a loop?
Because the agent is on your hardware, a runaway loop could consume your own billable compute or crash your internal dev server.
- Guardrail: Set resource limits (CPU/RAM) on the agent’s host container. Use the Cursor Dashboard to set a "Max Execution Time" for self-hosted tasks.
Is the code still sent to Cursor's servers?
No. The announcement explicitly states that "code and tool execution stay entirely in your own network." The Cursor control plane only sends the instructions and orchestration signals.
How do I handle secrets?
Do not hardcode secrets in your prompts. Since the agent runs in your infra, it should pull secrets from environment variables or a local vault (like HashiCorp Vault or AWS Secrets Manager) just like a human developer would.
What to do next
- Audit your compliance: Talk to your security team about the "Self-Hosted" architecture. Show them the isolated VM spec.
- Start small: Pick one repetitive internal task (e.g., "Update dependencies in all microservices") and let the self-hosted agent handle it.
- Connect your tools: Map your Slack or Linear triggers to the Cursor API to move from "manual coding" to "autonomous maintenance."