Hermes Agent Gains Vercel AI Gateway and Sandbox Backends: The Agent Stack Goes Plug-and-Play

On August 7, Vercel announced that Hermes Agent can now use AI Gateway as its inference layer and run every agent command inside a Vercel Sandbox microVM. Hermes is the open-source, MIT-licensed agent from Nous Research (227k GitHub stars at the time of writing), built around a self-improving loop: it creates skills from its own experience, refines them during use, and keeps persistent memory across sessions. The Vercel integration is significant not because Hermes needed another model provider, but because it completes the pattern of the agent as a portable program: the model layer and the execution layer are both now fully yours to choose.

What shipped, concretely#

Two independent switches, both opt-in, both configurable from the CLI:

AI Gateway as the inference layer. Hermes now appears in the AI Gateway setup wizard, and the picker pulls live model availability and pricing. Once configured, the agent can route through 200+ models on the gateway, with no markup on token cost, and every request lands in the AI Gateway dashboard alongside all your other usage and spend. For an agent whose whole pitch is provider-agnosticism, that removes the last reason to hardcode a single vendor: you get one dashboard for model choice, latency, and cost across whatever the agent touches.

Vercel Sandbox as a terminal backend. Hermes runs its shell commands locally until you set terminal.backend to vercel_sandbox. After that, each agent command executes in an isolated cloud microVM with a workspace root of /vercel/sandbox, instead of on your machine. The backend supports node24 (default), node22, and python3.13 runtimes. Local development authenticates with VERCEL_OIDC_TOKEN (a short-lived OIDC credential from vercel link and vercel env pull), not a long-lived API key.

For an existing install, the switch is three commands:

hermes update
hermes setup model      # pick Vercel AI Gateway
hermes setup terminal   # pick Vercel Sandbox
hermes doctor           # verify

This makes Vercel Sandbox the seventh terminal backend in Hermes' list, alongside local, Docker, SSH, Singularity, Modal, and Daytona. The agent has been a bring-your-own-execution tool for a while; Vercel is now one of the turnkey options rather than something you wire up by hand.

Why this matters to developers#

Three takeaways, in order of how much they change your setup.

Agents are becoming BYO infrastructure, and that is the healthy version of the market. Hermes already ran on models from Nous Portal, OpenRouter, OpenAI, and your own endpoints. AI Gateway slots in as the neutral middleman: one provider-agnostic API, one spend surface, 200+ models with no markup. If you are running several agents or several apps through the gateway, the spend-budget scoping Vercel added earlier this month now covers the agent's traffic too, with hard dollar limits that reject requests. Your cost guardrails stop being per-tool and become per-account.

Sandboxing is the feature, not a footnote. Running an agent's commands in a cloud microVM instead of your laptop is a real security posture change. Hermes is a general agent: it browses, schedules automations, runs subagents, and executes shell commands. Pointing that at your local filesystem is convenient and genuinely risky; an isolated microVM that exists for the duration of a command and can be revoked via OIDC is the containment layer most agent setups are missing. The cost is opt-in and cheap: microVMs that only spin up while the agent is working, with no idle baseline.

The agent control plane is consolidating, and both major clouds are racing there. Cloudflare unified Workers AI and AI Gateway into a single AI control plane on the same day this shipped. Vercel is positioning AI Gateway as the same thing from its side: the neutral routing, observability, and cost layer that sits between your agents and every model vendor. When two infrastructure vendors ship the identical abstraction in the same week, it is a signal that the winning position in the agent stack is not the agent itself, but the plane it runs on. For developers, the practical effect is that switching agents becomes cheaper than switching gateways, and switching models is now a config change in both.

The skills angle#

Hermes is also part of the skills ecosystem: it auto-generates skills from solved problems and supports the agentskills.io open standard. That puts it in the same design family as the skill compilation and typed harnesses wave, where agent capability is defined in portable files rather than baked into one vendor's runtime. The combination with Vercel's infrastructure is coherent: portable skills define what the agent can do, a neutral gateway defines what it can call, and a sandbox defines where it runs. None of the three layers cares which vendor the others come from.

Continue Reading#

• The Vercel Agentic Infrastructure Stack - how AI Gateway, Sandbox, and the agent runtime fit together as a platform
• Cloudflare Folds Workers AI Into AI Gateway - the same consolidation story from Cloudflare's side, shipped the same week
• The Agent Containment Capability Ledger - what isolation really covers, and where microVMs fit
• Vercel AI Gateway Adds Spend Budgets - the cost-cap math for agent workloads
• SIGIL: Skill Compilation With Typed Harnesses - why portable skills are becoming the standard unit of agent capability

Sources#

• Vercel Changelog: Vercel AI Gateway and Vercel Sandbox now available on Hermes Agent (fetched August 8, 2026)
• Hermes Agent by Nous Research (fetched August 8, 2026)
• NousResearch/hermes-agent on GitHub (repo stats fetched August 8, 2026)
• Vercel Docs: Configuring Hermes for AI Gateway
• Vercel Sandbox documentation