Security & Trust

AgentKit is a curated registry: maintainers select, vendor, and hash-pin each skill. This page is the honest version of what that buys you — and what it does not. Agents can read the same content as the MCP resource agentkit://docs/security.

What "curated" means #

  • Skills are selected by maintainers from public upstream repositories. There is no signup, no publish endpoint, and no per-account ownership.
  • Content is vendored: the skill's own directory (SKILL.md plus its references, assets, and examples) is copied into the registry and served from that pinned copy — never proxied or rewritten at request time.
  • Every entry publishes a provenance record: upstream repository, upstream commit, SHA-256 of its SKILL.md set, a combined SHA-256 over its supporting files, the bundle SHA-256, the SPDX license (or unknown), the publisher, and a modified flag.
  • Curated means "a maintainer looked at it and pinned it." It does not mean "audited for safety", and it does not mean AgentKit endorses what the instructions tell an agent to do.

Who can publish #

  • Only the maintainers, by adding a catalog entry and vendoring the upstream content. There is no self-serve publishing today.
  • To propose a skill, call submit_skill_feedback with kind: "idea", naming the upstream repository and commit. Proposals are persisted and reviewed out of band; the MCP tool surface for the catalog is read-only.
  • Entries are accepted when the upstream project is public and maintained, the skills are real instruction files, a license can be determined, and the content passes the risk review. When no license is detectable the entry says unknown instead of inventing one.

How review works #

  1. Read the vendored SKILL.md and every supporting file for instructions that would surprise a user.
  2. Run the risk scanner and record its flags with file and line evidence in the provenance record.
  3. Record the upstream commit and hashes, then publish the version.
  4. Re-scan when upstream changes — a content change always publishes a new version.

This is a human review of instructions plus an automated text scan. It is not a sandbox, not malware analysis of anything the skill ships, and not a runtime guarantee.

Commit pinning & hashes #

  • Versions are derived, never hand-written: the label is the upstream commit (7 characters) when the commit is pinned, otherwise unpinned-<bundle-hash8>. A label collision with different content appends build metadata (<label>+<hash8>), so one version can never mean two bundles.
  • get_skill_manifest and list_skill_versions publish the pinned hashes; diff_skill_versions compares two versions by content hash.
  • get_skill re-hashes the SKILL.md before serving it. If the bytes no longer match the pinned hash for the current version it refuses with content_hash_mismatch rather than serving unreviewed content under a reviewed label.
  • After fetching instructions, hash them yourself and compare against provenance.skillMdSha256.

Risk flags #

Every entry carries advisory flags with file and line evidence, shown in get_skill_manifest and in search_skills results.

  • pipe-to-shell — pipes a download straight into a shell (curl … | sh, wget … | bash)
  • shell — contains shell commands an agent may run
  • script — ships or invokes executable scripts (.sh, .py, .js)
  • destructive — rm -rf, force-push, DROP TABLE, mkfs
  • network — curl, wget, ssh, git clone, fetch(, requests.get(
  • file-access — reads local files or credential locations (.env, ~/.ssh, ~/.aws)
  • secrets — references API keys, tokens, or passwords

Flags are not verdicts. A flagged skill can be entirely legitimate (a deployment skill runs npm install; a security skill runs curl). No flags means the scanner found nothing, not that the skill is safe. Treat any flag as a prompt to read the evidence before running the instructions.

Report a malicious skill #

  1. Call submit_skill_feedback with kind: "issue", the skill name, and what you observed (expected vs actual). Include the version label so the report pins to exact content.
  2. If it is actively harmful, say so plainly and name the file and line.
  3. Maintainers triage reports. Takedown removes the entry, after which get_skill returns skill_not_found and the entry disappears from search_skills.
  4. Feedback is rate limited per IP (5/minute, 20/hour) and stores only a SHA-256 hash of the IP address — never the raw address, never an identity.

Known limitations #

  • Most vendored content predates commit pinning. Those entries report upstreamCommit: null, commitPinned: false, and modified: true with modifiedReason: "unknown-upstream": AgentKit cannot prove the bytes are unmodified upstream content. It does not mean the text was edited — it means it is unverified.
  • AgentKit ships a filtered subset of each upstream repository (skill files only, size-capped), so a vendored skill is never byte-identical to the whole upstream project.
  • There is no signature verification, no SBOM, and no reproducible-build attestation.
  • The scanner is regex-based text analysis: it misses obfuscated intent and dynamic execution, and it produces false positives.
  • Licenses come from the upstream repository's license metadata, resolved once. unknown means no detectable license; redistribution is the reader's responsibility.
  • No sandboxing. AgentKit serves text. Whatever an agent does with it runs in your environment with your permissions — show the user the instructions, hashes, and flags, and get confirmation before installing.
  • Review and re-vendoring are manual and periodic; upstream changes are not monitored continuously.
  • Some catalog entries are AgentKit-authored categories with no vendored content. They are not instruction sets, and their manifests and READMEs say so.