Google’s managed agents in the Gemini API can now run their tools on a developer’s own machines. Google AI Studio posted on X on Wednesday: “now you can bring your own sandbox with self-hosted environments”. Until now, the agents ran in Linux sandboxes hosted by Google.
Google’s documentation names eight places to run the self-hosted option. Six are outside companies: Cloudflare, Daytona, E2B, Fly.io, Modal and Vercel, by our count. Google’s own sandbox stays the default, and Google does not bill for its compute during the preview, so a self-hosted sandbox puts that compute on the developer’s side.
Business Pill · Least Privilege
An evergreen Business Engineer explainer on least privilege: giving an AI agent only the access its task needs. It teaches the general idea only and does not discuss Google’s release.
The key insight: As we read it, Google has split its managed agents in two: the agent loop and the model stay on Google’s side, billed at Gemini list rates, while the tools can now run in a sandbox the developer controls. Six of the eight places Google’s docs list for that sandbox are outside companies, by our count.
The documentation says a self-hosted environment “lets managed agents execute tools inside compute infrastructure you control”. The sandbox connects over outbound HTTPS using the open-source Gemini API CLI, and “requires no inbound firewall ports”.
Setup takes three steps, according to the docs. A developer creates an environment with the mode set to self-hosted and gets an ID and a one-time key, runs “gemini-api environments connect” inside the sandbox with that key, then passes the ID when starting an agent task.
The self-hosted guide on GitHub says the agent’s built-in file and shell tools, such as “execute_bash” and “grep_search”, run inside the developer’s infrastructure. The repository’s history shows the guides were added on 10 November.
The eight provider examples for self-hosted environments listed in Google’s managed-agent documentation, read on 12 November 2026: Cloudflare Sandbox, Daytona, E2B, Fly.io, Modal and Vercel Sandbox (outside companies, our count), plus Google Cloud Run and Agent Substrate. Google’s hosted sandbox remains the default; its compute is not billed during the preview.
The Eight Places It Runs
The documentation lists provider examples for Google Cloud Run, Agent Substrate, Cloudflare Sandbox, Daytona, E2B, Fly.io (Sprites and Machines), Modal and Vercel Sandbox. It adds that the runtime can run in Docker or rootless Podman, or locally for testing.
Local mode runs “execute_bash” directly on the host unless the developer passes a flag to disable it, the docs say. Daytona’s own guide says that command runs model-generated commands without the file tools’ confinement, “which is why the worker belongs in a sandbox rather than on your own machine”.
For Kubernetes and container clusters, the GitHub guide describes a split mode: a control-plane poller holds the API key, while the sandbox containers hold “zero API keys” and get only a short-lived lease token for one interaction.
Calendar days after managed agents launched in public preview in the Gemini API on 19 May 2026: the Antigravity agent 09-2026 (17 September, 121 days) and the Gemini API CLI beta (27 October, 161 days), from Google’s release notes, and the self-hosted environment guides committed to the gemini-api-cli repository on GitHub (10 November, 175 days). Day counts are our arithmetic.
What Changes for the Bill
Google’s pricing page says all model inference for managed agents is charged at standard Gemini list rates, and that “Environment compute (CPU, memory, sandbox execution) is not billed during the preview period”. The environment docs say the same of Google-hosted sandboxes.
Google’s docs cap the sources it mounts into a hosted sandbox: Git repositories at 500 MB, Cloud Storage sources at 2 GB and inline files at 2 MB in total. Self-hosted environments reject those hosted-only options for sources, network, environment variables and file snapshots with an error, according to the docs.
As we read the pricing page, the trade-off is cost and control: Google’s hosted sandbox is not billed during the preview, while a self-hosted one runs on compute the developer pays for, under rules the developer sets.
As we read the GitHub guide, the split mode is built so that the sandbox containers never hold the project’s API key: they get only a short-lived lease token for one interaction.
Gemini API docs, Environments in managed agents
“Instead of using the Google-managed sandbox, you can connect a self-hosted environment.”
Three Implications
FOR DEVELOPERS Self-hosted environments reject the hosted-only options for sources, network, environment variables and file snapshots with an error, the docs say.
FOR SANDBOX COMPANIESGoogle’s docs link a setup example for each of Cloudflare, Daytona, E2B, Fly.io, Modal and Vercel; Daytona has published its own guide.
FOR GOOGLE Model inference is charged at standard Gemini list rates wherever the tools run, according to the pricing page.
The framework says: “The surface, not the model, governs autonomy.”
As we read it, Google’s change moves one part of that surface, where the agent’s tools execute, onto infrastructure the developer chooses, while the agent loop and the model stay with Google. The framework predates this release and does not discuss it.
What Is Not Established
Google has not said in the documents we read when the preview ends or what it will charge for hosted environment compute after it. The docs give no price comparison between hosted and self-hosted sandboxes, and the outside providers’ own charges are not part of Google’s pricing page.
The provider list consists of setup examples in Google’s docs; apart from Daytona’s guide, which we read, we did not check whether each provider has announced the integration. Our count of six outside companies excludes Google Cloud Run and Agent Substrate.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Gemini API managed agents can now run their tools in a developer’s own sandbox. Google’s docs name eight places to run it, six of them outside companies by our count, while Google’s own hosted sandbox remains the default and its compute is not billed during the preview.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. We read Google’s managed-agent environment documentation, the agents section of the Gemini API pricing page and the Gemini API release notes on 12 November 2026, plus the self-hosted guide in Google’s open-source gemini-api-cli repository on GitHub and its commit history, Google AI Studio’s post on X of 11 November and the opening of Daytona’s integration guide. The count of outside companies and the day counts are ours. Google was not contacted for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
