eve/ The Agent Framework Workshop

Setup & Scaffold

Create the project, connect a model, and explore the filesystem that defines your agent.

Run the preflight first

Complete these checks before eve init. The initializer can guide you through model setup, but it cannot replace an outdated Vercel CLI while it is already running.

  1. Confirm Node.js and npm.

    node --version npm --version

    Use Node.js 24 or newer. npm is included with Node.js.

  2. Install or update the Vercel CLI.

    npm install -g vercel@latest

    Then confirm which version your shell will use:

    vercel --version

    If the command is missing or still reports the old version, open a new terminal and run which -a vercel. Multiple Node or package-manager installations can leave an older command earlier in your PATH.

  3. Confirm Vercel authentication.

    vercel whoami

    If you are not signed in, run:

    vercel login vercel whoami

    Make sure this account can create or access a project in the Vercel team you plan to use.

Recommended for this workshop: use AI Gateway through a Vercel project. The project supplies a short-lived VERCEL_OIDC_TOKEN, so there is no model API key to paste into the agent.

Create the agent

  1. Scaffold a new eve project.

    npx eve@latest init support-agent

    The command creates the directory, installs dependencies, initializes Git, and starts the local development experience.

  2. Enter the project.

    If the setup flow returned you to your shell, run:

    cd support-agent
  3. Connect a model.

    Your agent needs a language model to read its instructions, call tools, and reply in the terminal. A fresh scaffold opens the /model setup flow in the terminal UI. You will see three credential paths:

    PathCredentialWhen to use
    AI Gateway through a Vercel project — recommendedVERCEL_OIDC_TOKEN in .env.localChoose a team and project. eve links the project and pulls a short-lived token automatically.
    AI Gateway through an API keyAI_GATEWAY_API_KEYPaste a key from the Vercel dashboard when you do not want to link a project.
    OpenAI or Anthropic directlyOPENAI_API_KEY or ANTHROPIC_API_KEYFollow the printed provider instructions. This path may also require the matching AI SDK provider package and model configuration.

    You can reopen this picker at any time by typing /model.

  4. Send the first message.

    Prompt
    What information should a support engineer collect before triaging a customer issue?

    At this point the agent has only its starter instructions. The important milestone is that the model answers and the terminal streams the turn.

Understand the scaffold

Stop the dev server with Ctrl+C when you need the shell, then inspect the project:

support-agent/
├── agent/
│   ├── agent.ts
│   ├── instructions.md
│   └── channels/
│       └── eve.ts
├── evals/
├── package.json
└── tsconfig.json
PathPurpose
agent/instructions.mdAlways-on identity and rules
agent/agent.tsModel and runtime configuration
agent/tools/Typed actions the model can call
agent/skills/Procedures loaded only when relevant
agent/connections/MCP and OpenAPI services
agent/channels/Terminal, HTTP, Slack, web, or custom entry points
agent/subagents/Specialist child agents
agent/schedules/Recurring work
evals/Repeatable checks for the agent

The path supplies the capability name. For example, agent/tools/lookup_account.ts becomes the lookup_account tool without a separate registration step.

Restart and inspect

  1. Start local development.

    npm run dev

    eve dev runs the agent and opens an interactive terminal UI. Use /help to see its commands.

  2. Confirm what eve discovered.

    In a second terminal, from the project root:

    npx eve info

    This is the fastest check when an instruction, tool, skill, or channel does not appear.

You are ready when

Troubleshooting

Setup is stuck or did not finish

Press Ctrl+C and open the generated agent directory. If .vercel/project.json and .env.local already exist, the project was linked successfully—restart with npm run dev.

Otherwise, finish the setup manually:

npm install -g vercel@latest vercel whoami vercel link vercel env pull .env.local npm run dev

Keep .env.local private and out of Git. If the model warning remains after restarting, type /model and choose your provider again.