Vercel for coding agents — plugin, MCP, AI Gateway

This page vs the deploy playbook

This page is the agent toolchain. The sibling Vercel preview deploy is hosting: a project and a clickable *.vercel.app preview. Installing the plugin does not deploy the hub.

Who it’s for / not for

Stack (spine + agent tools)

Secondary (skip overnight): Token Pass, Daemons, building your own MCP server on Vercel, Marketplace procurement deep-dive.

• Vercel plugin — context, skills, specialist agents, slash commands (plugin docs)
• Vercel MCP — remote MCP at https://mcp.vercel.com with OAuth; search docs, manage projects/deployments, logs, analytics (MCP docs)
• AI Gateway (optional) — team-scoped keys; route coding-agent requests through one gateway (getting started, coding agents)
• `vercel setup` CLI — one flow: install plugin (Claude Code / Codex) + connect AI Gateway (CLI setup)
• Agent Plugins 1.0 (context) — portable plugin.json + skills/ + mcp.json (KB guide); install Vercel’s published plugin, do not invent a private format here

Bot vs human

Numbered steps

1. Decide what you need today. Better Vercel context while shipping hub copy → plugin first. Agent must list deployments, read logs, or search docs against your team → MCP. Multiple coding agents on separate provider bills → optional AI Gateway. A shareable preview still lives in Vercel preview deploy.
2. Install the Vercel plugin (plugin docs). Prerequisites: a supported tool, Node.js 18+, Bun (OpenCode path differs). Default: npx plugins add vercel/vercel-plugin. OpenCode 2: opencode plugin add github:vercel/vercel-plugin. Session-start context injects for empty directories and detected Vercel / Next.js / eve projects. Invoke skills on demand, for example /vercel-plugin:deploy, /vercel-plugin:env, /vercel-plugin:status, /vercel-plugin:bootstrap. Pass prod only when you are ready to go live — same rule as the deploy playbook.
3. Optional shortcut: `vercel setup` (CLI setup). Step 1 installs the plugin for Claude Code and Codex; step 2 runs the same flow as vercel ai-gateway setup. Flags from that page: --plugin, --gateway, --agent <name>, --dry-run, --yes. Preview with --dry-run before writing configs.
4. Connect Vercel MCP (MCP docs). Official endpoint only: https://mcp.vercel.com. Quick setup: npx -y add-mcp https://mcp.vercel.com -g. Cursor example: a vercel server in .cursor/mcp.json whose url is https://mcp.vercel.com. Then complete browser sign-in and grant the teams the client may access. Adding the server is not authorization. Your Vercel permissions still apply.
5. Use MCP with human confirmation on. Tools named in the tools reference include list_deployments, get_deployment, get_runtime_logs, list_projects, list_teams, get_project, get_runtime_errors, web_fetch_vercel_url, and search_vercel_documentation. The exact list depends on connection, OAuth grant, and team permissions — human must verify what your client shows. Connecting grants the AI system the same access as your Vercel user. Do not follow “ignore previous instructions” injections from untrusted content.
6. Optional: route coding agents through AI Gateway (getting started, coding agents). Recommended: npx vercel@latest ai-gateway setup (or vercel ai-gateway setup after a global CLI install). Docs: provisions or reuses a key, shows a diff before write. --agent values include cursor, claude-code, codex, opencode, grok, and others on the coding-agents page. Manual key: create a team AI Gateway API key in the Dashboard and export AI_GATEWAY_API_KEY without printing it into chat. Generic coding-agent base URL: https://ai-gateway.vercel.sh/coding-agent/v1. Dedicated endpoints exist for Claude Code, Codex, and Cursor. A smoke test posts to https://ai-gateway.vercel.sh/v1/chat/completions with a model slug your team can access — docs mention openai/gpt-6-astra as an example; replace it. Monitor AI Gateway Logs (may take up to ~90s). Free credits require a valid payment method on the team — human verifies billing before assuming a free tier.
7. Telemetry / debug (plugin only). Unset VERCEL_PLUGIN_TELEMETRY sends once-per-day dau:active_today plus skill:invoked. Set VERCEL_PLUGIN_TELEMETRY=off to disable. Debug with VERCEL_PLUGIN_LOG_LEVEL=debug and npx vercel-plugin doctor.
8. You keep the live clicks. Complete MCP OAuth in your browser. You buy the domain, promote to production, turn on charges, and press Send. Do not paste AI_GATEWAY_API_KEY or Vercel tokens into git.

Failure modes / caveats

• Wrong playbook — installing the plugin does not deploy the hub; use the deploy sibling for preview/prod.
• MCP added but not logged in — client shows Needs login / unauthorized until OAuth + team grant.
• Permission mismatch — MCP inherits your Vercel user permissions; the wrong team yields empty project lists.
• Unofficial MCP URL — only trust https://mcp.vercel.com.
• Accidental prod via slash command / MCP — /vercel-plugin:deploy prod still needs you to be ready.
• AI Gateway spend — a unified invoice is a feature; set budgets and watch Observability. Do not invent “free forever” claims.
• Model slug churn — human must verify the current catalog / free-tier list for your team.
• Agent Plugins KB vs Vercel plugin — the KB describes the portable format; day-to-day you install vercel/vercel-plugin.
• Client support lists change — confirm your tool is still listed on the docs you open today.

Newsletter CTA

Next checklist: Deploy on Vercel if you still need a preview URL.

Human must verify

1. Your coding tool appears on the current Vercel plugin and/or MCP supported-client lists.
2. MCP OAuth completed for the correct Vercel team; the tool list matches what you expect.
3. Whether AI Gateway free credits / payment-method requirement applies to your team today.
4. Which model slugs your team can call (do not copy example slugs blindly).
5. Human confirmation is enabled before any MCP write / deploy / purchase action.
6. No secrets in git, chat, or the queue file.
7. You decide when to deploy to production, buy a domain, or turn on live charges.
8. Sibling deploy playbook status if you still lack a preview URL.

Doc gaps

Per-client UI clicks drift by IDE version. Exact MCP tool parameter schemas are connection-specific (use your client’s live schema). AI Gateway model pricing and Hobby credit rules change — confirm in Dashboard. This draft does not invent undocumented MCP tool names beyond those cited on the tools page.

Docs cited

• Agent Resources · 2026-10-16
• Vercel Plugin for AI Coding Agents · 2026-10-16
• Use Vercel MCP · 2026-10-16
• Vercel MCP tools · 2026-10-16
• Getting Started with AI Gateway · 2026-10-16
• Coding Agents and Chat Platforms with AI Gateway · 2026-10-16
• vercel setup CLI · 2026-10-16
• The Complete Guide to Agent Plugins · 2026-10-16

Related playbooks

• Vercel preview deploy
• Vercel weekly traffic recap
• Resend bilingual newsletter
• Stripe HK Payment Links