Self-Hosted
Run benchmax in your own cloud.
Your traces, your rubrics, your benchmark — nothing leaves your network. One Docker command, or a Claude Code–guided install.
01·Single-command deploy
Docker Compose or Helm. Pick your poison.
One command stands up benchmax end-to-end. Compose for most teams, a Helm chart for Kubernetes shops.
~/benchmax
$ docker compose -f benchmax.yml up -d
[+] Running 6/6
✓benchmax-apistarted0.8s
✓benchmax-workerstarted1.1s
✓benchmax-webstarted0.6s
✓postgresstarted2.3s
✓redisstarted0.4s
✓grader-proxystarted0.9s
→ benchmax running at http://benchmax.internal
$
02·Claude Code setup
Tell Claude Code to install it
Point Claude Code at our install instructions and walk away. It provisions resources, installs the stack, configures grader access, hands you a running instance.
cc
Install plan● ready01Provision resources
EC2 t3.large · RDS db.t3.medium · VPC subnet
02Install stack
docker compose up · health checks green
03Configure grader
ANTHROPIC_API_KEY set · proxy routed
04Seed schema + admin user
[email protected] · temp password emailed
05Health check
all 6 services responding · latency < 50ms
→ running at https://benchmax.internal · 4m 22s total
03·Data stays put
Traces never leave your network
All traces, issues, rubrics, and test results live in your Postgres. Graders call the model provider you approve — OpenAI, Anthropic, Bedrock, or a self-hosted endpoint.
Your VPC · data residency
your network
API
node
Worker
node
Web
node
Postgres
all traces · issues · rubrics · runs
grader-proxyoutbound only
OpenAI
configurable
Anthropic
configurable
Bedrock
or self-host
Only grader API calls leave your network. Never trace data.