Eve agent costs: optimize Vercel or self-host
Eve deployments get expensive when sandbox compute stays provisioned, sandboxes multiply across subagents, or snapshots accumulate — not because a single model call ran long. Model tokens, AI Gateway, Vercel Functions, and Workflows bill separately; shrinking sandboxes does not remove those charges.
Next action: Estimate my sandbox cost with your monthly execution episodes and idle hours, or open the idle-time diagnostic if the agent already finished but the bill did not.
Upstream:
Eve bills through Functions, Workflows, Sandbox, and model routes — sandbox is only one line.
Eve pricingLast verified: 2026-09-10 (eve@0.52.5). Source record: docs/eve-sources.md in this repository.
Eve is a filesystem-first framework for durable backend agents on Vercel. Each agent has an isolated sandbox for bash and file tools. On Vercel, that sandbox is typically a Vercel Sandbox microVM; self-hosted deployments can use Docker, microsandbox, or other backends — with different cost and security profiles.
Simple cost breakdown
| Layer | What it is | Typical meter | Moves when you self-host execution? |
|---|---|---|---|
| Sandbox execution | Isolated /workspace compute for tools | Provisioned memory GB-hours, active CPU hours, creations, snapshots, egress | Yes — replace vercel() backend |
| Agent HTTP / Nitro | Routes under /eve/ | Functions or your own host | Partially — you operate the Node service |
| Workflows | Durable sessions, turns, streams | Workflow events and storage | Depends on workflow world |
| Models & tools | LLM and external APIs | Tokens + third-party usage | No — still external |
Do not treat “move off Vercel Sandbox” as eliminating inference or SaaS tool spend.
Three paths
1. Diagnose the bill
Your agent finished but spend continues → How to reduce Eve’s Vercel Sandbox bill.
Covers idle provisioned time, vCPU sizing, template reuse, subagent sandboxes, retries, and snapshot retention. Each section lists what to inspect, what to change, tradeoffs, and how to verify.
2. Estimate costs
Compare current, optimized, and self-hosted execution scenarios with transparent assumptions → Eve cost calculator.
The calculator uses published iad1 Pro sandbox rates (verified 2026-09-10). You enter execution episodes, runtime, idle time, and resources — not a full invoice upload.
3. Explore self-hosting
When managed sandbox stays costly after optimization → Self-host Eve: setup, costs, and tradeoffs.
Covers backend differences (Docker vs VM vs simulated shell), single-tenant recipes, operational responsibilities, and when staying managed is rational.
Decision guide
Guides in this cluster
| Page | Intent |
|---|---|
| Reduce sandbox costs | Operational fixes on Vercel |
| Cost calculator | Transparent Vercel vs self-host estimate |
| Self-hosting | Execution backend migration, not model hosting |
Related wiki content
- Vercel AI Gateway usage — model routing spend (separate from sandbox)
- Self-hosting hub — on-prem inference stacks (Ollama, vLLM) — not Eve agent runtime
- Where to start — general token optimization sequence