therandomsecurityguy/signet
AI EngineeringA graph-first registry that publishes, validates, and cryptographically certifies agent and skill versions for teams running autonomous agents.
What it does
Signet is a semantic governance registry for autonomous agents and computational skills. Instead of acting as a plain metadata catalog, it separates durable AGENT/SKILL identities from immutable version snapshots and moves each version through a lifecycle: draft → validated → published → certified, with rejected, deprecated, revoked, and archived as off-ramps. The goal is that both humans and AI agents can discover a capability, load its real artifact, inspect where it came from, and only use versions that passed policy checks.
Validation is layered: W3C SHACL graph-shape validation, ontology checks on capability claims, an advisory LLM semantic pass (non-blocking), and human review. Publication envelopes are canonicalized (RFC 8785 for JSON, RDFC-1.0 for RDF) and signed with detached signatures verified against trusted keys. Every lifecycle operation emits a governance event into an append-only Merkle log with inclusion proofs, and can anchor to an external Tessera-backed transparency log with checkpoints and witnesses. Provenance is modeled in PROV-O and can be replayed and verified locally.
It targets platform and security engineers who need to manage a growing internal pool of agent files, skills, and runbooks, and who want machine-checkable trust rather than undocumented trust.
Quick start
The fastest local path needs Node.js 22 and no Docker, AWS, or Kubernetes:
for d in service/api service/mcp service/webapp/knowledge-manager client/registry-cli; do
(cd "$d" && npm ci)
done
scripts/demo-local.sh demo
This starts three services against local JSON-file and filesystem backends and seeds a demo pack. Verify each is up:
curl -s http://localhost:3000/healthz
curl -s http://localhost:3100/healthz
curl -s http://localhost:3200/healthz
The MCP gateway runs on port 3000, the API on 3100, and the web console on 3200. A recorded 65-second walkthrough shows a SHACL policy being registered, a draft rejected with the policy message, and registry-operator verify publication returning verdict: verified recorded walkthrough. The same flow is documented step by step in the local demo guide, and Docker Compose and local kind variants keep the no-AWS property.
Usage notes
- Runtime: Node.js 22 is required. The quick path uses the
json_filegraph backend and filesystem CAS by default; Fuseki/SPARQL is optional for operators who want live graph inspection. - Bootstrap ordering: skills must be seeded before agents that reference them, and lifecycle-aware bootstrap should use the skill lifecycle endpoints rather than only
POST /v1/skills. - Optional components fail closed: advanced SHACL-AF rules, SPARQL functions, and custom SPARQL constraints require the optional pyshacl runtime sidecar; policies needing them fail closed without it.
- Auth: write routes on
service/apirequire authentication and persist audit state in SQLite. Mint API keys withservice/api'screate-api-keyscript. - Maturity: several areas are explicitly planned or evolving (richer ontology mining, fuller WORM/Trillian integration, witness-network semantics). Early-phase external warehouse and Okta connections may run in mock mode.
- License: BSD 3-Clause.
HotRepo's take
Signet is a very new repository, created in November 2026, and its only recorded ranked appearance is a single TypeScript new-project snapshot on 2026-11-18: 109 stars, 2 forks, and 5 stars gained over a one-hour observation window. That is too short a series to say anything about adoption or momentum, and it never entered the all-language daily rank. What stands out from the material is scope: within its first days it publishes a working lifecycle state machine, SHACL validation, dual-canonicalization signing, an internal Merkle governance log, and OpenAPI/SDK/MCP surfaces, which is a lot of trust machinery to implement so quickly — worth reading the design docs and tests rather than judging by the star count.
Written by AI from the project's README, updated . The project's own docs are the source of truth.