eve

eve

eve is a filesystem-first framework for durable backend agents on Vercel.

You author an agent as a directory on disk. The directory is the contract — markdown for the parts a human should read like a spec, TypeScript for the parts that benefit from real types and runtime behavior.

The framework is called eve. The published npm package is eve. The CLI binary is eve.

Preview Terms and Safeguards

eve is currently a preview and subject to the Vercel beta terms; the framework, APIs, documentation, and behavior may change before general availability.

As the deployer, it is your responsibility to ensure your agent complies with applicable laws.

You are responsible for configuring approval policies, tool restrictions, connection scopes, route/session authorization, sandbox controls, telemetry exports, and other safeguards appropriate for your use case.

Before using eve with non-public, sensitive, regulated, or production data, review which default tools, custom tools, MCP tools, shell/file/web tools, connected services, subagents, schedules, and external actions are available to the agent.

Require human approval or other safeguards for sensitive, irreversible, regulated, financial, healthcare, employment, housing, legal, safety-impacting, user-impacting, or external side-effecting actions.

Unless you configure stricter controls, eve agents may operate with permissive settings, including tool execution without human approval where approval is omitted and sandbox network egress that is not deny-all. Do not rely on model behavior alone to prevent sensitive or irreversible actions.

What eve Prioritizes

• Markdown-first authoring for instructions and procedures
• TypeScript where typed runtime behavior matters
• Durable message runs and follow-up turns
• Inspectable compiled artifacts under .eve/
• Per-agent sandbox with optional authored overrides
• A stable HTTP protocol built around immutable session IDs
• A runtime model that keeps channels, harnesses, and workflow execution separate

Authored Directory

my-agent/
├── package.json
├── tsconfig.json
└── agent/
    ├── agent.ts           # additive runtime config (model, name, build, compaction, …)
    ├── instructions.md    # always-on instructions prompt
    ├── tools/             # typed executable integrations
    ├── skills/            # optional named procedures the model can load on demand
    ├── hooks/             # lifecycle and stream-event subscribers
    ├── channels/          # message ingress and delivery (HTTP, Slack, …)
    ├── connections/       # external MCP server connections
    ├── sandbox/           # the agent's single sandbox (optional override)
    ├── workspace/         # files seeded into the sandbox on each session
    ├── subagents/         # specialist child agents (reuse `defineAgent`)
    ├── schedules/         # recurring jobs
    └── lib/               # shared authored code imported by other files

Authoring Helpers

Every authored directory has a typed helper. Import each from the matching subpath:

Helper	Subpath	Authored Location
defineAgent(...)	eve	agent.ts, subagents/<id>/agent.ts
defineInstructions(...)	eve/instructions	instructions.ts (or instructions.md)
defineTool(...), defineBashTool(...), defineReadFileTool(...), defineWriteFileTool(...), disableTool(...)	eve/tools	tools/<name>.ts
defineSkill(...), getSkill(...)	eve/skills	skills/<name>.ts (or skills/<name>.md)
defineHook(...)	eve/hooks	hooks/<slug>.ts
defineChannel(...), POST, GET	eve/channels	channels/<name>.ts
eveChannel(...), slackChannel(...), vercelOidc(...)	eve/channels/eve, /slack, /auth	reused from channels/<name>.ts
defineSandbox(...)	eve/sandbox	sandbox.ts (or sandbox/sandbox.ts)
defineSchedule(...)	eve/schedules	schedules/<name>.ts (or schedules/<name>.md)
defineEval(...), defineEvalConfig(...)	eve/evals	evals/<name>.eval.ts, evals/evals.config.ts

Runtime accessors live on the subpath that owns the concern:

• getSession() — current session, turn, auth, parent lineage (eve/context)
• getSandbox() — live sandbox handle for the current agent (eve/sandbox)
• getSkill(identifier) — handle for a named skill visible to the current agent (eve/skills)
• getContext(key), requireContext(key), hasContext(key), setContext(key), ensureContext(key, factory) — unified context helpers (eve/context)

The complete API reference, including types and lower-level runtime primitives, is in the TypeScript API Reference.

Tiny Example

agent/instructions.md

You are a weather-focused assistant. Be concise, accurate, and explicit when you use a tool.

agent/tools/get_weather.ts

import { defineTool } from "eve/tools";
import { z } from "zod";

export default defineTool({
  description: "Get the current weather for a city.",
  inputSchema: z.object({
    city: z.string(),
  }),
  async execute(input) {
    return {
      city: input.city,
      condition: "Sunny",
      temperatureF: 72,
    };
  },
});

agent/agent.ts

import { defineAgent } from "eve";

export default defineAgent({
  model: "openai/gpt-5.4-mini",
});

Quick Start

npx eve@latest init my-agent

eve init writes a new agent with eve's default model. Pass --model openai/gpt-5.5 to choose another AI Gateway model, --reasoning high to set a reasoning effort, or --channel-web-nextjs to add the Web Chat application. It installs dependencies, initializes Git, and starts the development server. When it finds a supported coding-agent REPL, the handoff menu can open that REPL instead or exit. Targeting an existing project directory (eve init .) adds the agent files and missing dependencies instead. It does not create a Vercel project or deploy the agent.

CLI commands:

• eve (including npx eve) — initialize the current directory, or start development in an eve project
• eve init <name> — create a new agent
• eve info — discovery results and compiled artifacts
• eve build — compile .eve/ and build the host output
• eve start — serve the built .output/ app
• eve dev — start the local runtime and REPL
• eve set [--model <model-id>] [--reasoning <effort>] — change root model settings
• eve extension init <name> — create a new extension package
• eve extension build — build an extension package

Deploying

eve is built to be durable. The runtime is Nitro + Workflows. Read the deployment guide for the deployment path, environment variables, and configuration.

Read Next

These files ship inside the installed package at node_modules/eve/docs/:

• Full docs index — recommended entry point
• Getting Started — install, scaffold, and run locally
• Project Layout — every authored directory in depth
• agent.ts — agent config reference
• TypeScript API Reference — complete define* and runtime helper reference
• Vercel Deployment — deploy to production

By authoring concern: Tools · Channels · Hooks · Skills · Sandbox · Connections · Subagents · Schedules · Evals

By runtime concern: Sessions and Streaming · Session Context · Context Control · Auth and Route Protection · CLI, Build, and Debugging · Instrumentation

Architecture (Internals)

You do not need this section to author an eve agent — it documents the public HTTP protocol contracts so eve composes predictably with other systems.

eve's internal split is:

• the channel normalizes inbound transport, applies auth and delivery policy, and owns channel-local addresses
• the harness does one unit of AI work and returns { session, next }
• the runtime persists state, follows next, streams events, and owns workflow primitives (start(), resumeHook(), createHook(), getWritable())

The public HTTP protocol exposes one immutable identifier: sessionId. Create a session explicitly, then use that ID for follow-up messages, controls, streaming, and inspection. Channel-local addresses remain behind authored channel APIs.

Changelog

See ./CHANGELOG.md for the release history. The changelog ships inside the published package so agents can read it directly from node_modules/eve/CHANGELOG.md to evaluate upgrades.