Reliable sandboxes
for AI benchmarks.

Harbor-native sandbox infrastructure that prevents agents from cheating.

terminal

$ harbor run \
    --dataset terminal-bench@2.0 \
    --agent claude-code \
    --model anthropic/claude-opus-4-6 \
    --env islo \
    --n-concurrent 500

Get started Read the docs

Why Islo?

Gateway profiles

Per-sandbox network rules. Control what agents can reach by host, path, method, and rate limit.

gateway.rule

host    api.github.com
path    /v1/*
action  allow
methods GET, POST
rate    100 req/min

Content filters

Scan response bodies for leaked answers. If an agent fetches a solution from GitHub, the gateway blocks it.

gateway.log

→ GET github.com/org/repo/issues/418
← 200 body received
  "def solve(puzzle): …"
✗ BLOCKED — content filter match

Credential injection

Agents get access, never keys. Secrets are injected by the proxy and stay out of trajectories.

inject.yaml

match   api.openai.com
header  Authorization
value   {{ secrets.OPENAI_KEY }}

	Islo	DIY	Daytona	Modal	Others
Snapshot-based environments Instant resets to a known state — no cold starts	✓	partial	✓	✓	✓
Gateway profiles Per-sandbox network policies with host/path rules, method filters, and rate limits	✓	—	—	—	—
Content filters Inspect and block request/response bodies — stop exfiltration in the payload	✓	—	—	—	—
Credential injection proxy Inject secrets without exposing them to agents	✓	—	—	—	—
Cost limits per run Cap spend per task or rollout — no runaway budgets	✓	—	—	—	—
Built-in trajectory storage Per-task records, queryable after runs	✓	—	—	—	—
Zero infra to maintain Fully managed — no orchestration to operate	✓	—	✓	✓	✓

Pricing

Pay for actual CPU cycles, resident memory, and storage. Nothing else.

CPU Time

$0.07/CPU-hour

Memory Time

$0.04/GB-hour

Storage Time

$0.0007/GB-hour

Example: 500 tasks on SWE-bench-verified (~10 min avg)

CPU

$23.33

Memory

$3.33

Storage

$0.12

Total / per task

~$26.78 / $0.054

Who we are

Proven at scale

We're the team behind Incredibuild — a decade of making parallel, compute-heavy workloads fast and reliable. Running thousands of sandboxes in parallel is core to what we do.

Built from real feedback

We've worked directly with the teams building evals for frontier labs. We understand exactly where existing tools fall short — and built Islo to close those gaps.

Fully committed

This is Incredibuild's strategic focus — not a side project. We're investing in the long-term infrastructure layer for AI evaluation, and we're here to support teams that depend on it.

Stop building infra. Start shipping evals.