Originally published on sandscale.dev.
A coding agent that can edit files and run commands needs two kinds of limits: a boundary around where it runs, and rules about what it may do inside that boundary. This post covers how Sandscale handles both for three agents people commonly use, Claude Code, Codex and Gemini CLI. It also says where the controls differ between them, because they do not all work the same way.
Everything here comes from sandscale.dev as of 20 November 2026. We only describe steps the site documents. Where a step is not documented, we say so instead of guessing.
Two ways to run an agent on Sandscale
1. The agent runs inside a Sandscale workspace. A workspace is a persistent cloud computer: its disk survives, you open it in the browser or over SSH, and it stops when idle. You run Claude Code, Codex, Gemini CLI or your own agent inside it, and it keeps working when your laptop is closed. Workspaces come with Pro and up. Hours count against your plan only while the workspace runs. See workspaces and the workspace docs.
2. The agent runs on your machine and uses Sandscale sandboxes over MCP. The agent connects to Sandscale's MCP server and calls tools to start a sandbox, run commands in it, move files and stop it. MCP works on every plan.
Connecting an agent over MCP
The documented command for Claude Code is:
claude mcp add --transport http sandscale https://mcp.sandscale.dev/mcp The first time the agent connects, it opens a Sandscale sign-in and then acts as you. For Codex, VS Code, Windsurf, Claude and ChatGPT, the quickstart says to add a remote MCP server with the same address (https://mcp.sandscale.dev/mcp). The docs show a Cursor configuration too, in ~/.cursor/mcp.json.
What the connection can do, from Connect Claude or ChatGPT: start, run, pause, snapshot and stop sandboxes; move files in and out; list, start and stop your workspaces; look at your object storage buckets; and list the names of your Vault secrets. What it cannot do: read Vault values (it lists names only), act as anyone but you, or go unrecorded. Every call goes into your organization's audit log with the tool, region, outcome and app, and never the arguments. Each app can make up to 120 calls a minute, and each person up to 60.
What policy does, per agent
This is the part to read carefully. Sandscale's agent policy page describes three steps:
- Write the policy, per project: what agents may do, what needs a person, and what is never allowed.
- Enforce it. Claude Code's hooks ask Sandscale before each tool call and prompt.
- See it. Every decision is in the session record and the audit log.
For Claude Code, that means enforcement: each prompt is checked, and each tool call is allowed, blocked or sent to a person. Per the page, Claude Code's hooks ask Sandscale before each tool call and prompt.
For Codex and Gemini CLI, the page says they "report every step", and the sandboxes page says they report over OpenTelemetry. The pages describe recording for these two agents; they do not describe blocking a call before it runs. If you need a hard block, rely on the isolation boundary, the network allowlist and the plan-level policies on templates, sizes and run times, and treat the per-call record as an after-the-fact trail.
What a policy can say
The documented scope of policies, from approvals and policies: which templates, sizes and run times members may use, and which agents may run. Approvals come with every plan; policies for projects come with Team and Enterprise. There is also an organization kill switch that stops its agents at once.
What happens when a call needs a person
When policy marks a call as needing sign-off, the agent pauses on it. A named approver approves or denies it with a reason in the console's approvals inbox, and nobody can approve their own request. An approved call runs once. The request and the decision are in the audit log. This works on every plan, though project policies that mark which calls need approval are on Team and Enterprise.
The rest of the boundary
Policy sits on top of the isolation boundary, not in place of it.
- Isolation. Each sandbox is a Firecracker microVM with its own kernel. For the general trade-offs, see our post on isolating agent-run code.
- Network. Agents that call model APIs need outbound access. Free uses an egress allowlist; Pro and up can open the network or keep an allowlist (see pricing).
- Secrets and DLP. Secrets live in Vault and do not go into prompts. DLP scans prompts and MCP tool calls before they leave, on every plan. See Vault and DLP.
- Record. Every action goes into a hash-chained audit log. All plans can read it; export is Team and up.
Handing tasks to agents: Agent Cloud
For coding tasks, Agent Cloud gives a task to an agent such as Claude Code or Codex running in one of your workspaces. You watch it work and can pause or stop it. The result is a branch and a pull request for you to review. Each step goes through your organization's policies and approvals and lands in the audit log. Agent Cloud comes with Pro and up, runs in workspaces, and uses your plan's vCPU-hours.
The review step matters: the output is a pull request, not a merge. A person reads the diff before it lands.
A sensible setup, using only what is documented
- Create a Pro (or higher) account and a workspace from a template (see the workspace guide).
- Run your agent in the workspace.
- Keep the network on an allowlist where you can.
- Put secrets in Vault, not in the repo or the prompt.
- For Claude Code, write a policy that marks the risky tool calls as needing a person, and choose approvers other than the person who started the run.
- Read the audit log after the first few runs and adjust.
- For Codex and Gemini CLI, treat the recorded steps as a trail and keep the boundary tight.
What this does not cover
- Agent behavior inside the box. The boundary and the policy limit what a call can do. They do not make the agent's decisions correct.
- Model-side behavior. "We provide the machine, not the model" is how the FAQ puts it: you bring the agent and its model access.
- Anything we have not documented. If this post says something is not published, that is where Sandscale's docs currently end. We would rather say so than fill the gap.
If you want to try it, pricing shows which plan includes workspaces and policies, and the quickstart has the MCP connection commands.

