Built by NPCs
Menu

Vercel eve deep dive: real deployments, what it costs on Vercel versus Google Cloud, and the alternatives

Week 4211 min read
Share
Save

Opens the assistant with a prompt. It asks for a short summary of this story, and for the assistant to remember that you saved it.

TL;DR

eve is Vercel's Apache-2.0 agent framework, in public preview since 17 June 2026. Vercel runs 100+ agents on it, and an outside company reports about 20 on one 4-core server. In an illustrative model, Vercel hosting costs less than a Google Cloud VM until roughly 13,000 to 26,000 runs a month.

On this page

eve is Vercel's Apache-2.0 agent framework, in public preview since 17 June 2026. Vercel says it runs more than 100 agents on it, an outside builder runs a fleet of ten, and one company reports about 20 eve agents on a single 4-core Linux server. In our illustrative model, hosting on Vercel costs less than a Google Cloud VM until roughly 13,000 to 26,000 runs a month.

vercel/eve: The Framework for Building Agents (GitHub)The repository. README, docs and the self-hosting discussion thread.github.com →
At a glance
Licence
Apache-2.0
Status
Public preview since 17 June 2026 under Vercel's beta terms. The README warns that APIs may change before general availability
Maintained by
Vercel. One copy of the repo page listed 53 contributors
Activity
The copy read showed release eve@0.27.0 from 21 July 2026, 3.9k stars and 369 forks. Other cached copies differ, so treat counts as indicative
Install
npx eve@latest init my-agent, which needs Node.js 24 or newer

What eve is

Vercel calls eve Next.js for agents. An agent is a directory. agent/instructions.md is the system prompt, agent/tools/ holds one TypeScript file per tool with the filename as the tool name, agent/skills/ holds markdown playbooks the model loads when relevant, and channels/ and schedules/ hold the Slack, Discord or cron entry points. The smallest working agent is two files, a model and a set of instructions.

The framework compiles that directory and supplies the production pieces most teams hand-roll:

  • Durable sessions. Each conversation runs as a checkpointed workflow built on the open-source Workflow SDK, so it can pause, survive a deploy or crash, and resume.
  • A sandbox per agent. Agent-written code runs in an isolated environment, on Vercel Sandbox when hosted and on Docker, microsandbox or just-bash locally.
  • Approvals. A tool can require a person to approve a call, and the session waits without using compute. Vercel's docs also show how to auto-approve routine tool calls with Jev, the model from our earlier story, and ask a human only for risky ones.
  • Connections, channels, subagents, schedules, tracing and evals, each as a file or a command.

Models are named as strings such as openai/gpt-6-astra, routed through Vercel's AI Gateway, or passed as a provider object from the AI SDK for direct provider access.

Who runs it for real

The evidence is thin and mostly self-reported. Vercel's figures are its own, and the two outside accounts below come from the people running the agents. No independent audit of any of them was found.

Who What they run Reported result Evidence
Vercel More than 100 agents in production, including d0, a data analyst in Slack, Lead Agent, a sales rep, Athena, a pipeline cockpit, and Vertex, a support agent d0 handles over 30,000 questions a month. Lead Agent costs about $5,000 a year to run and returns 32 times that, with one engineer part-time. Vertex solves 92% of tickets on its own. Athena was built in six weeks and pipeline coverage nearly doubled Vendor's launch post, no outside check
Zack Proser, Applied AI at WorkOS Nine single-purpose bots plus a Chief of Staff orchestrator on his website, and a meta-agent that opens PRs against the others from Slack Each bot runs its own durable session, hands work over authenticated POSTs and parks outbound actions for his approval His own review. He reports lost hours on setup and version drift
A company on GitHub About 20 eve agents on one 4-core, 16 GB Linux server inside the company network, for over two months, with systemd, a bubblewrap sandbox and shared Postgres workflows The poster built an open-source operations layer, Eveland, to manage releases, gateways and dashboards Maintainer's post. The company is unnamed

Around them sit public templates and integrations. A Browser Use template runs a web-browsing agent, a Railway template deploys agents from the open EveAgents directory, and Langfuse, Raindrop and Arcjet publish eve integrations. Vercel's own self-hosted reference is a single-host Docker setup, which it describes as a baseline that makes no claim to high availability.

What self-hosting costs on Google Cloud

eve's self-hosting guide says eve build then eve start produces a standard Node server you can run in a container, a VM or behind your own proxy. It gives no cost figures, so the numbers below are estimates built from published Google Cloud rates in the us-central1 region. Prices change, so check them before budgeting.

Setup Machine Compute per month Notes
One agent or a small team, one VM e2-standard-2, 2 vCPU and 8 GB About $49 on demand, about $31 with a one-year commitment, about $21 to $24 on spot Spot machines can be reclaimed, so they suit dev and non-critical agents. Source: Holori's price table
A fleet on one box e2-standard-4, 4 vCPU and 16 GB About $98, twice the e2-standard-2 rate My arithmetic. This matches the 4-core, 16 GB server the company in the GitHub thread runs about 20 agents on, though not on Google Cloud
Shared workflow state Cloud SQL for PostgreSQL About $8 for a shared-core instance, about $100 for 2 vCPU and 7.5 GB The second figure is my arithmetic from May 2026 per-hour rates of $0.0413 per vCPU, $0.007 per GB of memory and $0.22 per GB of SSD. Source: Usage.ai and Bytebase

By default eve keeps workflow state on local disk under .eve/.workflow-data, so a single instance needs only a persistent disk and no database. A shared Postgres workflow world is for several instances or a fleet that must share state, and the docs mark that setting as experimental.

The model bill will dominate. As an illustration, take a run that reads 150,000 tokens and writes 15,000, on a model priced like Gemini 4 Argon at $4 and $20 per million tokens. That run costs about $0.90, or about $0.45 at Argon's introductory $2 and $10 rate through November. A hundred such runs a day cost about $90 a day, roughly $2,700 a month against about $98 for the server. Vercel's own figure for its sales agent, about $5,000 a year all-in, is the vendor's number and gives a sense of scale for one busy agent.

What the same workload costs on Vercel

On Vercel, eve is billed by usage. The Pro plan costs $20 per seat per month and includes $20 of usage credit, and the Hobby plan is free for personal, non-commercial projects. Eve's pricing page says its costs come from Functions, Workflows, Sandbox and AI Gateway, at these published rates:

Meter Rate
Functions Active CPU $0.128 per hour, memory $0.0106 per GB-hour, invocations $0.60 per million. Time spent waiting on a model is not billed as CPU
Workflows $0.02 per 1,000 events, $0.50 per GB written and $0.50 per GB-month retained. A normal step produces at least three events
Sandbox Active CPU $0.128 per hour, memory $0.0212 per GB-hour, $0.60 per million creations, storage $0.08 per GB-month
AI Gateway The provider's list price with no markup. Team accounts get $5 of free credit a month until they buy credits

Because the gateway adds no markup, model tokens cost the same on either host, and only the hosting layer differs. To compare, take an illustrative workload. Each run has 20 steps, 10 seconds of active CPU and 1 GB of memory held for 5 minutes. One run in three uses a 2 vCPU, 4 GB sandbox for 2 minutes with 30 seconds of active CPU. At 3,000 runs a month my arithmetic gives:

  • Workflow events: 180,000 events, about $3.60.
  • Function CPU and memory: about $1.07 and $2.65.
  • Sandbox CPU and memory: about $1.07 and $2.83.

That is about $11 of usage, which the Pro credit absorbs, so the bill is the $20 seat. Usage grows at roughly $0.0037 a run in this model.

Runs per month Vercel Pro Google Cloud
3,000 About $20, with usage inside the credit About $49 for an always-on e2-standard-2
30,000 About $112 About $98 for an e2-standard-4, if one 4-core box carries the load, plus $8 to $100 if you add Cloud SQL

In this model Vercel stays cheaper until roughly 13,000 runs a month against the small VM and 26,000 against the 4-core box. Past that, a flat-priced server wins on raw infrastructure cost. The comparison leaves out your time. On Google Cloud you patch, monitor and secure the box, run the sandbox and database, and give up the Vercel-only extras listed below. The workload shape is my assumption, so real agents with longer sandbox sessions or heavy event counts will shift the crossover.

What you give up, and the catches

A reply in the GitHub self-hosting thread says the core runtime behaves the same off Vercel, and lists what you lose: the Agent Runs dashboard, Vercel Cron, sandbox prewarm, latest-deployment routing, and domain-level network allow-lists for sandboxes. The Docker sandbox backend is allow-all or deny-all only. For traces you use the OpenTelemetry exporters, and Langfuse and others document eve setups.

The catches that matter in practice:

  • Two proxy routes are needed. Forward both /eve/ and /.well-known/workflow/ without rewriting paths. If only /eve/ is proxied, a session starts but the run stalls because workflow callbacks never arrive.
  • Do not rely on Vercel sign-in. The guide says to avoid using vercelOidc() as the only production authenticator outside Vercel. Configure Basic auth, JWT, OIDC or a custom verifier.
  • Version pinning. The workflow layer is on a 5.0.0-beta line and the runtime rejects incompatible protocol versions. Zack Proser's review also advises pinning eve and its AI SDK and Connect packages and committing the lockfile.
  • One host is a single point of failure. Vercel's own self-hosted reference says it is a baseline without high availability, and that its Docker backend does not expose CPU and memory quotas.
  • Fit with Google Cloud services. The default sandbox picks Docker, microsandbox or just-bash by what is available, so it points to a VM or a Kubernetes node where Docker runs. That is my reading, so check it before choosing a serverless container service.
  • Model access. Set AI_GATEWAY_API_KEY to keep using string model IDs from off Vercel, or install a provider package and pass a model object with that provider's key.

The alternatives, and who suggests them

None of the builders quoted calls any alternative simply better than eve. The advice from builders is conditional on where the agent will run. Two voices stand out, and neither framed eve as inferior.

Alternative Who suggests it Why The catch
Mastra Zack Proser says it is the alternative that comes up most when eve's Vercel tie is a problem. Astro co-creator Fred Schott lists it among the "OG agent frameworks" TypeScript, Apache-2.0, agents, tools and workflows with suspend and resume, memory, and portability across clouds You stand up and operate the durability, sandbox and connector layer that eve gives you on Vercel, per Proser
Flue Fred Schott, who tells the Latent Space podcast that eve is the most directly competitive framework Apache-2.0 TypeScript framework built around a programmable agent runtime, deploying to Node, Cloudflare, GitHub Actions and GitLab CI It is flagged experimental and was 1.0 beta in June. Developers noted its API resembles eve's, and its durable state leans on Cloudflare Durable Objects
Vercel AI SDK Vercel positions it as the lower-level layer under eve Version 7 adds durable workflow agents, approvals and telemetry, per Developers Digest You build channels, sandbox and scheduling yourself
OpenAI Agents SDK, Claude Agent SDK Framework comparisons from Speakeasy and Cipher Projects Thin primitives and the tightest fit with one provider's features Provider lock-in. The OpenAI SDK ships no sandbox or channels
LangGraph Comparison guides for Python-heavy orchestration Graph-based workflows with state checkpoints Bring your own sandbox

Cipher Projects also advises comparing durable runners such as Temporal and Inngest separately from agent frameworks, since eve's durability comes from Vercel Workflows.

Best for

  • TypeScript teams already on Vercel who want durable agents, Slack channels, approvals and a sandbox without building the plumbing, and who can accept a preview-status framework. Proser's pattern of many small single-purpose agents is the use it handles best.

Skip if

  • leaving Vercel is a hard requirement, you need a stable production platform this quarter without pinning and piloting, you want to learn agent internals from the ground up, or you plan to run on a serverless container service that cannot run Docker.

Sources

  1. eve product page, Introducing eve (Vercel blog, 17 June) and changelog: what eve is, the launch date and Vercel's own agents.
  2. eve getting started, self-hosting guide and Vercel's eve docs: requirements, self-hosting steps, Jev approvals, beta status. Eve pricing and limits: how usage maps to Vercel resources.
  3. Workflow pricing, Sandbox pricing, AI Gateway pricing, Pro plan and Vercel's pricing page: the Vercel rates.
  4. vercel/eve on GitHub and discussion 83: licence, activity, what self-hosting loses, the 20-agent server. vercel-labs/steve: the single-host reference.
  5. Zack Proser's review: a nine-bot fleet, sharp edges and the Mastra alternative.
  6. Latent Space on Flue, Firecrawl on Flue and eve, The Road to Enterprise and Tiny Agents: Flue, its maker's view of eve and how the two compare.
  7. Speakeasy, Cipher Projects (OpenAI SDK), Cipher Projects (durability) and Developers Digest: framework comparisons.
  8. Browser Use template, Railway template and Langfuse: ecosystem examples.
  9. Holori e2-standard-2, Usage.ai on Cloud SQL and Bytebase on Cloud SQL: Google Cloud prices.

Continue reading