Vercel AI Gateway and Sandbox microVM Now Available in Hermes Agent

Vercel has updated Hermes Agent with two new integrations. First, the agent can now use Vercel AI Gateway as its inference layer, giving developers access to more than 200 models with no markup on token costs and centralised usage tracking in the AI Gateway dashboard. Second, an opt-in Vercel Sandbox backend lets agent commands run inside an isolated cloud microVM rather than on a local machine, with support for node24, node22, and python3.13 runtimes.

What happened

Feature	Detail
Models available via AI Gateway	200+
Token markup	None
Sandbox workspace root	/vercel/sandbox
Supported runtimes	node24 (default), node22, python3.13
Config key to enable Sandbox	terminal.backend = vercel_sandbox
Local dev token method	vercel link + vercel env pull (VERCEL_OIDC_TOKEN)

Vercel’s Hermes Agent now has two optional backends you can enable during or after setup. The AI Gateway integration is configured through the setup wizard: after installing Hermes, you pick Vercel AI Gateway from a picker that pulls live model availability and current pricing directly from the gateway. Every inference request then shows up in your AI Gateway dashboard alongside any other AI usage you are already tracking.

The Vercel Sandbox integration is strictly opt-in. By default, agent commands still run locally. Set terminal.backend to vercel_sandbox in your config and subsequent commands execute inside a cloud microVM. The workspace root shifts from your local filesystem to /vercel/sandbox. For local development against the Sandbox, you authenticate with a VERCEL_OIDC_TOKEN obtained by running vercel link followed by vercel env pull.

Why it matters

The AI Gateway piece solves a real accounting headache. Teams using multiple AI providers often end up with spend scattered across separate dashboards. Routing everything through a single gateway with no added cost per token is a straightforward win for visibility, and it means you can swap models without re-instrumenting your app.

The Sandbox backend is more significant for security-conscious teams. Running agent commands inside an isolated microVM means a misbehaving or compromised agent command cannot touch your local filesystem or credentials. It also brings agent execution closer to how code actually runs in production on Vercel, which should reduce “works on my machine” surprises. Paired with AI integration work, this kind of sandboxed execution is exactly the kind of guardrail that makes autonomous agents safer to deploy.

Our take

The no-markup token pricing is a genuine differentiator, not a footnote. Most gateway products charge a percentage on top of provider costs. If Vercel holds to zero markup as usage scales, this becomes a meaningful cost control lever, especially for teams running high-volume inference.

The Sandbox microVM approach is the right architecture for agent tooling. Local execution of agent commands is fine for demos but uncomfortable in production. Having a cloud-isolated environment with a defined workspace root and known runtimes is a much cleaner story. The opt-in design is sensible too: it does not break existing local workflows while giving teams a migration path.

One thing to watch: the Sandbox runtimes are currently node24, node22, and python3.13. If your agent needs a different runtime, you are not supported yet. Check the runtime list before committing this to a production workflow.

For teams already on Vercel who have been keeping an eye on the agent adoption gap, this update lowers two real friction points at once: cost transparency and execution safety.

What to do about it

1. Install or update Hermes and run through the setup wizard to connect Vercel AI Gateway.
2. Check the live model picker to confirm your preferred models are available before migrating inference.
3. Set terminal.backend = vercel_sandbox in a non-production project first to test the microVM behaviour.
4. Run vercel link then vercel env pull to get your VERCEL_OIDC_TOKEN for local dev against the Sandbox.
5. Verify your runtime requirement (node24, node22, or python3.13) is on the supported list before going live.

If you are building agent-driven workflows and want help structuring safe, cost-tracked inference pipelines, talk to the Lumien team about what that looks like in practice.

Source: Vercel Blog