See every agent run on Vercel
Agent Runs shows each eve session in the Vercel dashboard, with its turns, tool calls, reasoning, timing, and token usage. It works automatically for eve projects, with no instrumentation required.
Every step of every turn
Timings for each step, including skill loads and tool calls, alongside the turn's input, output, and reasoning.From your terminal or coding agent
Inspect runs with vercel agent-runs, or with the Agent Runs tools in Vercel MCP.Tokens and triggers
Runs over time by trigger, like Slack and HTTP, with input, cached, and output tokens for each turn.
Agents that run your company, power your product, or are personal
Start from a working template, or build your own from scratch.
Build a Slack agent that answers your team's business questions from your documented dbt models.
- Answers in SlackOnly for the people and channels you allow.
- Grounded in dbtReads your model docs before it writes a query.
- Read-only SQLQueries DuckDB or MotherDuck, and never writes to them.
Build an agent that turns issues into reviewed draft pull requests, from triage and planning to implementation.
- GitHub and LinearPicks up tasks where your team already files them.
- Four stationsClassifier, analyst, implementer, and reviewer subagents work each task.
- Draft pull requestsEvery task ends in a reviewed draft on your repository.
Agents inside your product, over web, voice, or iMessage, with memory for every user.
- Web chatInside your app, with long-term memory for each user.
- VoiceCalls that run tools and ask for approval.
- iMessageThreads that share memory with web and Slack.
Deploy your first agent in less than 60 seconds on Vercel
Define an agent as a directory of TypeScript and Markdown, and deploy it with one command.
Instructions and skills
Define the agent in Markdown, with skills it loads only when needed.Any model
Pick a model for each agent through AI Gateway, and switch with one line.
Own your intelligence
Customize at every layer, from the harness to the UI to every agent run.
- agent/tools
- bashDefault
- query_crmCustom
- sleepDisabled
- web_searchDefault
The agent's tools: bash on by default, a custom query_crm tool added, sleep disabled, and web_search on by default. Own your harness(opens in new tab)
Start with great defaults. Turn off what you don't need, extend what you do. Scale from side project to enterprise.
Three agent apps in overlapping windows, built with Next.js in front, Nuxt, and SvelteKit. Own your agent UI(opens in new tab)
Build rich agent UIs natively in Next.js, Nuxt, SvelteKit, and TanStack. Ship agent and app in one deploy.
Rounded squares represent data in your runtime, with model providers above and a hosting provider below. Choose your model routing and hosting infrastructure. Own your data(opens in new tab)
Control where your agent runs and who can access it. Bring your own workflow storage, sandbox provider, and model routing.
Everything you need to ship to production
One agent at the center of radar rings, reached from WhatsApp, GitHub, Slack, Linear, Discord, Telegram, Google Chat, Microsoft Teams, iMessage, Web. Agent chat on every channel(opens in new tab)
Multiplayer and single-player chat across Slack, Teams, iMessage, WhatsApp, and the web. Channels are a framework primitive.
- Resolved per user
- AliceMemoryCRM tools
- BobMemorySchedules
- CarolMemoryLegal skills
One deployment resolves capabilities per person in the Acme org, each avatar badged with the Acme logo: Alice gets memory and CRM tools, Bob memory and schedules, and Carol memory and legal skills. Multi-tenant with dynamic capabilities(opens in new tab)
One deployment, every customer. Skills, schedules, memory, and tools resolved per user at runtime.
Three runs: a fully agentic loop of model steps, a deterministic workflow of code steps, and a workflow with one agentic step among its code steps. Determinism where you need it(opens in new tab)
Fully agentic loops, deterministic workflows, or both in one run. You choose per task.
Any model, no rewrite(opens in new tab)
Swap models with a one-line edit. Provider fallback included. Route different models to different tasks.
Durable execution & scheduling(opens in new tab)
Checkpoints every step. Survives crashes, restarts, and days-long waits. Runs on crons, too.
Human in the loop(opens in new tab)
Pause for approval. Resume durably when it arrives.
Evals(opens in new tab)
Test agent behavior before and after every change.
Connections & subagents(opens in new tab)
Connect to services over MCP and OpenAPI, or delegate to focused subagents. Credentials stay in the runtime.
Agent identity & sandboxes(opens in new tab)
Bring any OIDC provider and act as the agent or the signed-in user. Code runs in an isolated sandbox per session.
Uses Vercel AI primitives
AI Gateway for model calls, Sandboxes, Drives for workspaces that persist across Sandbox runs, Workflows, and Connect. All critical agent infrastructure works out of the box, no gluing point solutions together.
