First-party Agent Skills for the Autonomi network — installable, self-contained instructions that let an AI agent read from, store on, build on and run the network from its own tools.
| Skill | What it does | Status |
|---|---|---|
autonomi |
Read data by content address; store files publicly or privately and get a permanent address back; run nodes that contribute spare disk and earn ANT; wire the network into an application. One skill, routed by task. The agent never sees a wallet key. | Prototype 0.1.4 — being tested with the community |
The skill follows the Agent Skills format, so it isn't tied to one tool. Pick whichever route your agent uses.
npx skills add WithAutonomi/skillsUseful flags: -a claude, -a codex, -a opencode or -a '*' to choose the agent; -g to install globally; -y to skip prompts. Update later with npx skills update autonomi.
/plugin marketplace add WithAutonomi/skills
/plugin install autonomi@withautonomi
Custom Claude Code marketplaces do not auto-update by default. Enable auto-update for withautonomi in /plugin → Marketplaces, or update explicitly with:
/plugin update autonomi@withautonomi
Future only. The earlier version-pinned installer metadata was removed because OpenClaw's current contract cannot express the release route without going stale. No working OpenClaw installer or public ClawHub listing is claimed.
Copy skills/autonomi/ — SKILL.md, VERSION and references/ — into wherever your agent loads skills from (for Claude Code, ~/.claude/skills/autonomi/). Keep the folder together: the references are loaded on demand. Copies installed by hand do not update automatically; repeat the installation to replace them with a newer release.
After an update, start a new agent session before relying on the new instructions. In Claude Code, /reload-plugins can activate an updated plugin without restarting when the client says a reload is available.
The agent loads only the skill component in skills/autonomi/. skills.sh and manual installations copy that directory; Claude Code caches the repository-root plugin package declared in .claude-plugin/marketplace.json, then discovers skills/autonomi/ within it as the skill component. Repo-side files are not loaded as skill instructions. On first use the agent detects or installs the upstream ant command-line client from its official GitHub releases — fetching and reading the installer before running it, or taking a checksum-verified manual path — and learns the tool from ant --help. The script installers do not yet verify checksums or signatures themselves, so universally verified delivery remains a target rather than a current guarantee. Nothing in the skill holds keys or moves funds. A paid write uses a SECRET_KEY the person provisions to the tool's environment themselves, or the person runs the paid command; the agent only ever works with public addresses, and spending is quote-show-wait by default.
Sandboxes. Installing
antneedsgithub.comand its release hosts reachable. The installer's version lookup usesapi.github.com, which some agent sandboxes block while allowing the download itself; the skill then falls back to a manual path that reads the version from the release checksum file. The network is peer-to-peer over UDP, so a proxy-only sandbox can install the tool but will seefound 0 peers— the skill says so rather than retrying. Distributing the CLI through npm, which every sandbox allows, is tracked in ant-client #190.
A prototype, deliberately: one skill for readers, writers, builders and node operators, to find out whether they can share a skill without any of them feeling the others' weight. What it has to prove, how it's tested and the evidence so far are in planning/TESTING.md; where every shipped claim comes from is in source-bindings/autonomi.md; current state and open threads are in planning/HANDOFF.md. The design record and ADRs under docs/ were written for the earlier node-operator-only skill; docs/DESIGN.md opens with a note on how they relate to what ships now.
Platform status: no platform has completed the full live prototype test. The Windows install path is source-read but has not been run on Windows. Treat its instructions as unverified until that test is complete; see
planning/HANDOFF.md.
Heads-up — binary install in locked-down sandboxes. The
antinstaller downloads its binary from GitHub's release CDN (release-assets.githubusercontent.com), which some AI-agent sandboxes block even whengithub.comis allowed. The skill detects this and tells you exactly what to allowlist rather than failing silently. Tracked as an upstream/release item inplanning/release-endpoint-accessibility.md.
skills/autonomi/ # skill component; copied directly by skills.sh and manual installs
SKILL.md # entry: what Autonomi is, ground rules, task router, keys & money, verified-against, further reading
VERSION # release version, kept in sync with skill and plugin metadata
references/ # on demand: install-and-verify, wallet-and-tokens, run-nodes, build-on-autonomi
.claude-plugin/ # Claude Code marketplace + plugin manifests (repo root is the plugin root)
docs/ # repo-side, never loaded as skill instructions: design, ADRs, archive
planning/ # current state, test protocol, briefs, parked threads
source-bindings/ # provenance for every shipped claim
scripts/ # ADR governance check (runs in CI)
Only skills/<name>/ is discovered as an agent skill. Claude Code may cache the repository-root plugin package, but repo-side files remain maintainer material rather than skill instructions.
Branch + PR, never direct to main; architecture, protocol and security decisions go through an ADR in docs/adr/. See CONTRIBUTING.md. Security reports: .github/SECURITY.md.
MIT or Apache-2.0, at your option — see LICENSE-MIT and LICENSE-APACHE.
Built by the Autonomi team (MaidSafe). Autonomi: https://autonomi.com · documentation: https://docs.autonomi.com · for developers: https://developers.autonomi.com