Open-source Agent Skills that teach AI coding agents to operate hiloop — isolated agent sandboxes with branchable, versioned workspaces and tree-native observability.
The skills follow the open Agent Skills standard, so they work across harnesses (Claude Code, Cursor,
Codex, and others). They drive hiloop through the hiloop CLI — the supported agent interface —
not an MCP server: a CLI the agent already knows how to drive costs far less context than loading tool
definitions every turn.
| Skill | Use it to |
|---|---|
autoresearch |
Run an autonomous research loop with evolving idea cards, scored experiments, an ensemble, and a leaderboard |
authenticating |
Sign in with hiloop login (or a key), verify identity, manage tenant scope and keys |
creating-sandboxes |
Create / inspect / delete sandboxes; pick a profile or image, request resources |
running-commands-in-a-sandbox |
Run commands (buffered, one-shot, or over SSH); move files in/out |
persisting-and-branching-workspaces |
Seal a workspace into a revision on stop, resume it exactly, branch N sandboxes from one revision |
assembling-a-personal-devbox |
A long-lived, owner-only devbox: managed SSH, rsync, access grants, suspend-and-wake |
managing-secrets |
Give a run a credential it uses but never sees (the secret broker) |
coordinating-with-leases |
Serialize concurrent agents with named, TTL-bounded leases (at most one live holder per name) |
launching-as-workloads |
Launch runs/sandboxes as a registered machine identity (a workload) and control who may launch as it |
querying-observability-trees |
Capture a run and query (SQL) / tail / diff its run-lineage telemetry |
annotating-runs |
Stamp structured judgments (outcome / score) you can filter and aggregate on |
reporting-product-bugs |
Report a hiloop bug (or send product feedback) to the hiloop team with hiloop feedback |
AGENTS.md is the whole-product orientation an agent reads first.
1. Install the CLI (single static binary):
curl -fsSL https://hiloop.ai/install.sh | sh
hiloop --version2. Authenticate — hiloop login is the default (--device on a remote box with no local browser):
hiloop login
hiloop whoami…or headless (an agent / CI), skip the browser with a key: export HILOOP_API_KEY="hil_…".
3. Install the skills. Choose your harness, or install every supported target:
hiloop skills install claude-code # cursor | codex | gemini | copilot
hiloop skills install allcopilot installs into the current repository; the other targets install for your user. Full target
paths, native alternatives, and unsupported-harness fallback instructions are in
SETUP.md.
Then ask your agent to spin up a hiloop sandbox, run work in it, and query the trace tree — the skills guide the rest.
For autonomous metric optimization, point the agent at your task, fixed dataset/scorer, and the
autoresearch skill. Watch it with the public
fleet-dashboard.
These skills mirror the hiloop CLI surface and the published guides at https://docs.hiloop.ai. Keep
skill bodies under ~500 lines and push dense schemas into references/, per the Agent Skills best
practices. When the CLI changes, update the matching skill in the same change.
scripts/check-skills.sh is a lightweight drift guard (run in CI on every change and weekly): it
checks each skill's frontmatter/length/links and — when a hiloop binary is on PATH — that every
hiloop … command a skill mentions still exists, derived live from hiloop --help (no hardcoded
list). Run it locally with ./scripts/check-skills.sh.
Apache-2.0 (see LICENSE).