An open-source TypeScript framework for building AI agents, released by Vercel as a public preview on June 17, 2026. Positioned as “what Next.js did for the web, for agents,” it defines an agent as a single directory. Instructions and skills are written in Markdown, tools in TypeScript, and the pieces you always end up needing in production — durable execution, sandboxed execution, and observability — come built in. It is licensed under Apache 2.0, with the source published on GitHub.
In agent development, calling a model is the easy part. The heavy lifting comes afterward: how to resume a session that died mid-run, where to safely execute generated code, and how to trace what actually happened. eve provides that groundwork as a set of conventions, so developers can focus on the agent itself.
Key Features
- Filesystem-first structure: An agent is a directory. Put files in fixed locations —
agent.ts(model configuration),instructions.md(system prompt),tools/(TypeScript functions),skills/(Markdown procedures),channels/(integrations),schedules/(recurring runs), andsubagents/(delegation) — and eve discovers them automatically and wires them into the agent loop. No configuration files or registration boilerplate required - Durable execution: Sessions are checkpointed at each step, so work continues across process crashes and deploys. A long-running agent execution will not vanish every time you redeploy
- Sandboxed execution: Code generated by the agent runs in an environment isolated from your application runtime — Docker locally, Vercel Sandbox in production
- Observability out of the box: Every run emits an OpenTelemetry trace covering model calls, tool invocations, and sandbox commands. Traces export to Braintrust, Honeycomb, Datadog, Jaeger, and others; on Vercel they surface in the Agent Runs dashboard
- One agent, many channels: The same agent can be exposed through channel adapters to Slack, Discord, Teams, Telegram, GitHub, Linear, and HTTP APIs. Adding a surface is a command such as
eve channels add slack - MCP and OpenAPI connections, plus human approval: External services connect via MCP servers or OpenAPI documents, with OAuth supported through Vercel Connect. Human-in-the-loop approval before execution, delegation to subagents, and evals for verifying behavior are all included
Pricing
| Item | Price | Details |
|---|---|---|
| eve (the framework) | Free | Apache 2.0 open source; can run on your own infrastructure |
| Vercel Hobby | $0 | For personal use. One developer seat |
| Vercel Pro | $20/month | Includes $20 of usage credit. Additional developer seats are $20/month each |
| Vercel Enterprise | Contact sales | SLAs, advanced security, dedicated support, and more |
Pricing is current as of August 2026. The framework itself is free, but hosting costs (if you deploy to Vercel) and the API usage fees of whichever AI model you use are billed separately. Check the official site for the latest pricing.
Pros & Cons
✅ Pros
- Durable execution, sandboxing, and observability — the things you inevitably need later in production — are there from day one
- The conventional directory layout makes someone else’s agent easy to read and reason about
- Skills are plain Markdown, so non-engineers can contribute knowledge in the form of written procedures
- The same agent can reach Slack, GitHub, and other surfaces without a separate implementation per channel
- Apache 2.0 open source, so you are less locked into a single hosting provider
⚠️ Cons
- It has been a public preview since June 2026, so the API and specification may still change
- Requires working knowledge of TypeScript and Node.js; this is not a no-code tool
- Vercel is the first deployment target to be fully supported; other platforms are still ahead
- Model API fees and hosting costs are separate, so it is not free end to end
- A convention-heavy design can feel constraining if you want to decide every structural detail yourself
Comparison with Similar Services
| Criteria | eve | Mastra | LangGraph | OpenAI Agents SDK |
|---|---|---|---|---|
| Provider | Vercel | Mastra | LangChain | OpenAI |
| Language | TypeScript | TypeScript | Python / TypeScript | Python / TypeScript |
| Structuring approach | Auto-discovery via directory conventions | Assembled in code | Described as a graph (state machine) | Assembled in code |
| Durable execution | Built in | Supported | Supported | Left to the implementation |
| Sandboxed execution | Built in (Docker / Vercel Sandbox) | Configured separately | Configured separately | Configured separately |
| Channel deployment | Slack, Discord, GitHub, and more supported natively | Implemented separately | Implemented separately | Implemented separately |
| License | Apache 2.0 | Open source | Open source | Open source |
Who Is It For
- TypeScript developers who want to take a prototype agent to something that survives real operation
- Teams that want one implementation of an internal Slack bot or GitHub-connected agent to serve multiple channels
- Anyone who wants to follow agent behavior through traces and iterate on it
- Existing Vercel users who want to deploy agents through the same workflow they already use
- Teams that want to bring procedures and knowledge in as Markdown and maintain them together with non-engineers
Summary
eve takes the foundation that agent projects tend to rebuild every time — durable execution, sandboxing, observability, channel integration — and absorbs it into framework conventions. “Drop a file in the directory and it works” keeps the learning curve low, and Markdown skills suit team-based operation. That said, it is still a public preview and assumes TypeScript knowledge, so a realistic approach is to build one small agent with npx eve@latest init, see how traces and day-to-day operation feel, and then decide on fuller adoption.