Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

5.3 KB
Raw Blame

Skill supply-chain scanner

Every marketplace skill is scanned before it can be listed, and again on every new version. A skill that tries to read secrets, call hosts it never declared, evaluate hidden code or talk the agent out of asking its owner is blocked or held for a moderator. The authors see why, with the file, the line and how to fix it.

Code: api/_lib/skill-supply-scan.js (scanner), api/_lib/skill-scan-store.js (policy, moderation, catalog rescan). Corpus and tests: tests/skill-supply-scan.corpus.js, tests/skill-supply-scan.test.js.

What it reads

handlers.js is parsed into an AST with acorn and walked; tools.json, SKILL.md and manifest.json are read as data and prose. Strings built at runtime are folded back to their value first (concatenation, template literals, String.fromCharCode, atob, Buffer.from(..., 'base64' | 'hex'), join, reversed splits), so globalThis['pro' + 'cess'] is seen as process.

Rule Severity Catches
env_access, secret_material, credential_store block process.env, key material on a wallet or keypair object, localStorage, cookies
secret_to_network, secret_with_network block (review when every host is declared and verified) a secret read that flows into, or sits beside, a network call
undeclared_host, dynamic_network_target review a call to a host manifest.json does not declare, or to a URL the scanner cannot resolve
exfil_sink block webhook catchers, raw IP hosts, the cloud metadata endpoint
dynamic_code, dynamic_import, forbidden_import, unverified_import block eval, new Function, .constructor(...), string timers, computed import(), Node built-ins, remote imports
obfuscated_string, obfuscated_identifiers, hidden_payload, encoded_blob, minified block code built to hide what it does
excess_permission, purpose_mismatch review spend, sign or message tools called without being declared in requires
secret_param, tool_text_injection block a tool that asks the model for a private key or seed phrase, or carries injected instructions
prose, skip_confirmation, hidden_text block or review SKILL.md that asks for secrets, tells the agent to act without asking the owner, hides instructions in comments or invisible characters

Safe phrasing is not flagged: "Never send funds without asking the owner to confirm" and "Read-only queries may execute without confirmation" both pass.

Verdicts

Verdict Effect
clean Lists normally.
review Saved privately with review_status: pending. A moderator approves or rejects with a note.
blocked Refused with 422 skill_blocked and the findings. Cannot be listed or approved. The only way out is an edit that scans clean.

A moderator approval holds across rescans of unchanged content. Any edit to the content, tools, manifest or handlers is a new version and is scanned again.

Declaring network hosts

Add the hosts a skill calls to manifest.json. Wildcards match subdomains.

{ "requires": { "network": ["api.weather.example.com", "*.cdn.example.com"] } }

A call to a declared host passes. A call to anything else is held for review; a call to a webhook catcher, raw IP or metadata endpoint is blocked whatever is declared.

API

Endpoint Purpose
POST /api/skills Publish. Accepts optional manifest and handlers_src. Returns scan in the 201, or 422 skill_blocked with scan.
PATCH /api/skills/:id Edit. Same body fields. Setting is_public: true on a held skill returns 409 held_for_review.
POST /api/skills/scan Dry run for the publish form. Body { name, description, content?, schema_json?, manifest?, handlers_src? }. Writes nothing. Signed-in, rate limited.
GET /api/skills/moderation Admin only. The pending queue with findings.
POST /api/skills/moderation Admin only. { id, decision: "approve" | "reject", note? }.

The scan payload is { verdict, counts, scanner_version, held, findings: [{ rule, severity, file, line, column, message, fix, snippet? }] }.

Suggested skills published from an agent (skill-candidates.md) go through the same scan.

In the publish form

Marketplace, Publish a skill. Open "Skill code and instructions", paste SKILL.md, manifest.json and handlers.js, and press Check skill to see the findings before publishing. A blocked publish shows the same list. A held publish is saved privately with a note that a moderator will list it.

Rescanning the catalog

node --env-file=.env.local scripts/rescan-marketplace-skills.mjs            # dry run
node --env-file=.env.local scripts/rescan-marketplace-skills.mjs --apply    # write verdicts, hide held or blocked
node --env-file=.env.local scripts/rescan-marketplace-skills.mjs --only-stale

The dry run prints how many skills would be hidden before anything changes. --only-stale rescans only rows scanned by an older scanner version. On 2026-10-24 the existing catalog of 152 skills was scanned: 152 clean, 0 review, 0 blocked.