Vercel announced eve at its Ship London conference on June 17, an open-source framework that takes a different approach to how agents are defined: every agent is a directory of files. The framework is Apache-2.0 licensed, TypeScript-first, and ships with first-party infrastructure including durable execution, sandboxed compute, and an AI Gateway, according to a detailed analysis by andrew.ooo.

The GitHub repository is live under the vercel organization.

Agents as Filesystem Artifacts

Most agent frameworks follow a similar pattern: import a library, instantiate an agent class, register tools as decorated functions, deploy as a long-running process. The “agent” exists at runtime with no canonical file to point at.

eve inverts this. A minimal eve agent is a directory containing an instructions.md system prompt, a tools/ folder where each TypeScript file defines one tool, and a skills/ folder of Markdown playbooks. Run eve dev and the framework reads, validates, and runs the directory as an agent.

This has practical consequences for teams. Agent changes become file diffs reviewable in pull requests. A team can see “this agent now has access to the delete-customer tool” as a one-line change, not buried in a module. Skills are portable Markdown files that can be copied between agents or forked from open-source repositories.

Built-in Infrastructure

eve does not just define a layout. It ships a runtime with several integrated capabilities, according to andrew.ooo:

  • Durable execution: Agent runs survive restarts, retries, and long tool calls via Vercel Functions checkpointing.
  • Sandboxes: Each tool call can run in isolated compute for code execution, browser interaction, and file I/O.
  • AI Gateway: A single endpoint routing to OpenAI, Anthropic, Google, and OpenRouter. Model swaps are a config change.
  • Subagents: Nested agents declared as subdirectories. A parent agent passes a task and yields.
  • Evals: Test cases defined as TypeScript files in agent/evals/ that run in CI.
  • Channels: Built-in adapters for HTTP, Slack, GitHub, webhooks, and CLI. One agent serves multiple surfaces.
  • Approvals: Tools can require human approval, surfaced through any channel.

Passport and Connect

Alongside eve, Vercel released two enterprise products. Vercel Connect (public beta) replaces static API keys with short-lived OAuth/OIDC credentials, giving tools task-scoped tokens to services like Slack, GitHub, Snowflake, and Salesforce. Vercel Passport puts internal apps and agents behind a corporate identity provider (Okta, Microsoft Entra) by default, with audit logs and central policy.

The identity products position Vercel against “shadow AI,” where engineers deploy internal agents that bypass corporate identity governance.

Where eve Fits

The closest existing comparison is OpenClaw, which already uses a skills/ directory pattern and treats agents as composable configurations. eve is more aggressive about making the entire agent a filesystem layout, and it bundles managed infrastructure that OpenClaw delegates to the host.

LangGraph offers stronger graph flexibility for cyclic state machines and complex branching. CrewAI and AutoGen use code-defined roles and handoffs. eve’s differentiator is that the agent is the directory: diff-able, reviewable, and deployable without writing orchestration code.

The framework is TypeScript-only. Python support is not on the immediate roadmap, which limits adoption among teams that have standardized on Python-based agent stacks.

Filesystem-as-Agent Convergence

eve’s launch adds to a pattern. Anthropic’s Claude Skills, OpenClaw’s skills directories, and now Vercel eve all converge on the same idea: agents defined as file structures rather than runtime code. For teams evaluating agent frameworks in mid-2026, the question is shifting from “which orchestration library” to “which deployment infrastructure,” and Vercel is betting that the answer is the same platform already hosting their web applications.