Eve
Eve is Vercel's open framework for building production agents, billed as "Next.js for agents." An agent is a directory: you define its instructions and skills in Markdown, write tools in TypeScript, and Eve compiles the directory into durable workflows with channels, sandboxes, and schedules. First-party sources: eve.dev, github.com/vercel/eve (Apache-2.0, about 4,750 stars on 2026-08-22), and the Vercel page at vercel.com/eve. Eve leans on Vercel's AI primitives: AI Gateway for model calls, Sandbox for isolated compute, Workflows for checkpointed execution, and Connect for authenticated MCP endpoints.
Compare OpenAI Agents SDK or LangGraph if you prefer a code-first runtime over a file-system convention.
Key Features
- An agent is a directory:
instructions.mddefines identity,skills/holds Markdown playbooks loaded when relevant,tools/files become callable tools by filename,channels/connects Slack, Discord, Teams, web chat, and API, andschedules/run agents on cron. - Durable by default: every step is checkpointed; agents park between messages and resume on delivery. Workflows survive crashes and restarts.
- Isolated sandbox: each agent runs with file tools and bash in an isolated VM, with optional
sandbox/sandbox.tsto choose a backend. - Subagents: delegate specialized work to child agents with their own prompts, tools, and sandboxes.
- Human-in-the-loop: tools that need confirmation trigger approval gates; sessions park until resolved.
- Evaluations: define test suites with scoring rubrics and run them on deployment or on a schedule.
Use Cases
- Teams shipping production agents that need to appear in Slack, Discord, web chat, or an API without gluing point solutions together.
- Scheduled automation: daily reports, weekly digests, and recurring jobs via cron schedules.
- Developers who want Markdown-first control over agent behavior and TypeScript tools instead of a visual builder.
Problems It Solves
- Infrastructure glue: model calls, sandboxes, persistence, and channels work out of the box.
- Lost work on crash: durable checkpointing means long agent runs survive restarts.
- Tool registration boilerplate: the filename becomes the tool name, no registration required.
Pricing
| Plan | Price | Notes |
|---|---|---|
| Open source | $0 | Apache-2.0, self-hostable via npx eve@latest init my-agent |
| Vercel hosting | Varies | Usage of AI Gateway, Sandbox, Workflows, Connect billed by Vercel |
Getting Started
- Run
npx eve@latest init my-agent. - Write
instructions.mddescribing the agent's role. - Add
tools/TypeScript files; the model can call them immediately. - Optionally add
skills/, asandbox/sandbox.ts, channel files, and cron schedules. - Deploy on Vercel or self-host; run evals before shipping.
Frequently Asked Questions
Is Eve free?
The framework is open source under Apache-2.0. Running it on Vercel uses your Vercel plan's AI Gateway, Sandbox, and Workflow quotas.
Do I need to learn a new DSL?
No. Instructions and skills are Markdown; tools are plain TypeScript with a defineTool helper.
Can Eve connect to Slack or Discord?
Yes, channel files make the same agent available in Slack, Discord, Teams, Google Chat, WhatsApp, web chat, and API.
Alternatives
- OpenAI Agents SDK: OpenAI's lightweight agent orchestration.
- LangGraph: graph-based stateful agent runtime.
Tips
- Keep
instructions.mdshort and move detail intoskills/so prompts stay focused. - Use subagents for tasks with separate toolsets instead of one large tool list.
- Add a cron schedule for recurring jobs and let Workflows handle restart durability.
Conclusion
Eve turns agent building into the same shaped experience as building a Next.js app: a directory, Markdown, and TypeScript. If you already live in the Vercel ecosystem and want durable multi-channel agents without infrastructure glue, start at eve.dev.
Comments
No comments yet. Be the first to comment!
Related Tools
DeepSeek Harness
deepseek.com/harness
MIT agent runtime. Rechecked: 144,198 GitHub stars. Pay model API only. Live Flash peak $0.44/$1.32 per 1M, not the old $0.14/$0.28 row.
Hermes Agent
hermes-agent.nousresearch.com
Self-improving open-source AI agent by Nous Research with a built-in learning loop, multi-channel gateway, and model-agnostic runtime. MIT licensed.
nanobot
github.com/HKUDS/nanobot
Ultra-lightweight, self-hosted personal AI agent framework in Python with WebUI, tools, long-term memory, MCP integrations, multi-agent workflows, and scheduled automation.
Related Insights
Seven AI Coding CLIs, Six Months: No Matter How Strong the Model, Work Needs Supervision
Claude Code, Codex, opencode, pi, omp and DeepSeek Harness all have personalities. After six months of deep use I run a division of labor: pi for the fastest cheapest reviews, omp for complex PRs, DeepSeek Harness on V4 Flash for high-frequency low-cost review, and Claude Code, Qoder and Cursor for writing. No matter how strong the model, work needs supervision — ideally from an independent third party.

Anthropic Subagent: The Multi-Agent Architecture Revolution
Deep dive into Anthropic multi-agent architecture design. Learn how Subagents break through context window limitations, achieve 90% performance improvements, and real-world applications in Claude Code.