Cerver — Infrastructure for production agents.

01 — Not all agents are the same

Five agents. Five different computers.

Support chat

machine: slim
model: GLM
time: under 1 s

Cheap

Video edit

machine: your GPU box
model: Claude
time: minutes

Heavy

Code fix

machine: E2B
model: Claude · plan
time: 20 min

Strong

Lead triage

machine: Vercel
model: gpt-5-mini
time: 2 s

Fast

Nightly research

machine: your laptop
model: Claude · plan
time: overnight

Free

Know what each one costs. Then cut it.

Support chatSonnet → GLM

$41

Video edit

$58

Code fixAPI → plan

$22

Lead triage

$9

Nightly research

$0

this week $130Example figures

02 — Start small

Start with one model call. Grow into agents.

# change one line in the app you have
ANTHROPIC_BASE_URL=https://gateway.cerver.ai/v2/proxy/anthropic
ANTHROPIC_API_KEY=ck_…

cost per callspend capkey stays server-side

# one chat per customer
POST /v2/sessions  {"metadata":{"external_user_id":"9b2c…"}}
POST /v2/sessions/:id/run-llm  {"input":"Where is my order?"}

transcriptscost per customerswap models mid-chat

# a real agent, on a real machine
cerver run --cli claude --on e2b "fix the failing test"
cerver run --cli claude --bill sub "…"   # on your Claude plan, $0

any CLIany machineyour subscription

# split a big job across cheap models
cerver split "research 30 companies, one page each"
cerver crons --project sales   # run it every night

parallel workersper-step costschedules

03 — The platform

Nine pieces you’d otherwise build.

Run

Sessions

Transcript, model, machine, cost, owner.

Compute

Laptop, server, E2B, Vercel. Move mid-run.

Crons

Scheduled runs that don't need your laptop awake.

Connect

Model proxy

Your vendor key never enters the sandbox.

Vault

Model keys per project.

MCP server

npx @cerver/mcp

Control

Caps

Per project. Blocked before the vendor sees it.

Routing

Easy turns to cheap models.

Redaction

Secrets scrubbed from prompts.

04 — Any CLI, any machine

Move a run. Keep every word.

here now

Your laptop

Claude Code on your Claude plan

$0 per token

here now

Your server

A Mac mini or a box you own, always on

$0 per token

here now

E2B

A fresh sandbox for each run

your key, at cost

here now

Vercel Sandbox

Next to the app it serves

your key, at cost

session 7f3a turn 1 transcript 1 turn, kept spent so far $0.0000

Animated example.

05 — Separate projects

A $50 cap stops at $50.

Project
support-chat

own keyown provider keys

$128 of $200 this month

doneNorthwind · claude-sonnet$0.31

runningGlobex · gpt-5-mini$0.02

Project
nightly-agents

own keyMac mini

Hit its $50 cap. New runs blocked.

donelead-research · claude (plan)$0

blockedlead-research · gpt-5over cap

Example figures.

06 — See every step

One model failed. The log said which.

replay at 5× speed · 82.7 s

Step	Model	Machine	Time	Cost
plan	claude-sonnet	MacBook (plan)	25.6 s	$0
email 1	gpt-5-mini	hosted	6.0 s	$0.00035
email 2	glm-4.5-flash	hosted	36.1 s	$0.00200
email 3	grok-4.3	hosted	3.4 s	failed: out of credits
email 4	gpt-5-mini	hosted	5.2 s	$0.00033
email 5	glm-4.5-flash	hosted	30.6 s	$0.00169
summary	glm-4.5-flash	hosted	21.0 s	$0.00125
total			82.7 s	$0.0056

Real session log, Nov 4.

07 — The test

Same job, two ways. The cheaper way lost.

Time

19.9 s vs 35.6 s

Extra cost

$0 vs $0.0023

Score out of 10

9 vs 6

Claude on a Claude plan vs. split across per-token models. One run, Nov 4.

08 — Compared

AgentCore if you live in AWS. Cerver if you don’t.

	AWS Bedrock AgentCore	Cerver
Where agents run	AWS only	Laptop, server, E2B, Vercel
Your Claude or ChatGPT plan	No	Yes, $0 per token
Spend caps	Rate limits only	Per project and per CLI
Separate customers	Isolated sessions	Cost per customer, built in
How you pay	Per second, per call	Monthly plan
Browser, evaluations, agent payments	Yes	No
AWS integration and compliance	Yes	No

Source: aws.amazon.com/bedrock/agentcore, Nov 2026.

09 — Pricing

Plans start at $89 a month.

Per-second platforms

It depends

CPU and memory per second
A fee per tool call
Tokens on top

Cerver Team

from $89 / month

Up to 10 developers
Tokens at cost, or $0 on your own plan
Your compute, or a sandbox at cost
Larger teams: talk to us

10 — Start

Install in one line. Free for one developer.

$ curl -fsSL https://cerver.ai/install.sh | bash

Already in Claude Code? Type /plugin marketplace add eyal-gor/cerver-claude-plugin. No card needed.