compare

Eve runs the agent. Something still has to remember.

Vercel Eve gives agents durable execution, sandboxing, and deployment infrastructure. gcontext gives them a persistent state folder they can read and write between sessions. These are complementary concerns.

What Vercel Eve is

Eve is an open-source, filesystem-first framework for building durable backend AI agents. Vercel released it in June 2026. Each agent is a directory of files: instructions in Markdown, tools in TypeScript, configuration for model providers, authentication, channels, and schedules.

The framework includes durable execution, sandboxed compute, approval workflows, tracing, and evals. It runs locally, on Vercel, or on any long-running Node.js host. It is Apache-2.0 licensed and published as the eve npm package.

What Eve is good at

  • Durable execution. Every conversation is a workflow. Each step is checkpointed. A session can pause for human approval and resume where it stopped.
  • Sandboxed compute. Code that agents generate runs in isolated environments.
  • Channels. Agents connect to Slack, email, webhooks, or custom interfaces.
  • Deployment. A single command deploys agents to Vercel infrastructure with observability and model routing.

These are execution concerns. Eve answers: how does the agent run, where does it run, and how do you operate it at scale?

Runtime vs state

Every agent system separates into two layers. The runtime layer handles execution: picking the next action, calling tools, recovering from failures. The state layer handles memory: what the agent knows, what it learned, what it decided.

Eve is a runtime. Session state persists through durable workflows. Long-term knowledge across sessions is a different problem. As of August 2026, Eve tracks first-class cross-session memory as a planned feature.

MCP-served state alongside Eve

Eve supports MCP servers as tool providers. An Eve agent can connect to a gcontext instance and get six tools: read_file, write_file, list_dir, grep, run_script, and agent.

The state folder lives outside Eve. You can inspect it with any text editor, version it with git, and back it up with standard file tools. If you move to a different runtime, the state folder comes along unchanged.

terminal
$ uv tool install gcontext-ai
Installed gcontext-ai

Write, then read next session

The Eve agent writes what it learned to the state folder. On the next run, it reads prior decisions before it acts.

The state folder is plain text. No database, no embeddings. Open any file and see exactly what the agent knows.

eve agent
$ gcontext write_file modules/onboarding/decisions.md \
"## 2026-08-22\n- Selected Postgres over DynamoDB"
wrote 2 lines to modules/onboarding/decisions.md
$ gcontext read_file modules/onboarding/decisions.md
## 2026-08-22
- Selected Postgres over DynamoDB

Comparison

Vercel Evegcontext
Agent executionRuns the agent, picks actions, calls toolsNot an execution layer
SchedulingCron schedules, event triggers, channelsNo scheduling (use your runtime)
Session persistenceDurable workflows checkpoint every stepNot involved in session state
Knowledge persistencePlanned feature (as of August 2026)Plain text files served over MCP
Cross-tool memoryWithin the session context windowFiles accessible from any MCP client
InspectionTracing and observability dashboardOpen the folder in any text editor or git

Questions

Is gcontext a competitor to Vercel Eve?

No. They solve different problems. Eve is a framework that runs agents in production with durable execution, sandboxing, and deployment infrastructure. gcontext is an MCP server that stores what agents know as plain text files. One runs the agent, the other holds its memory.

Does Eve persist agent memory?

Eve persists session state through durable workflows, so a conversation survives crashes and deploys. Long-term memory across sessions is not built in as of August 2026. The Eve team tracks first-class cross-session memory as a planned feature. Check Eve's documentation for current memory capabilities.

Related

What is agent state?
Definition and management patterns
AI agent memory
Four approaches compared
Stateful MCP servers
The pattern behind gcontext
Add persistent state to your agents

Install gcontext. Your runtime handles execution. The state folder handles memory.

View on GitHub