Starlight Intelligence System
frankxai/Starlight-Intelligence-System
Memória persistente e governança unificada para múltiplas ferramentas de IA
Ver no GitHubUm servidor MCP que fornece substrato de inteligência soberana com memória persistente, 48 agentes, 71 skills auto-ativadas e governança atestada para toda sua frota de ferramentas de IA (Claude, Cursor, Codex, Gemini CLI). Sincroniza contexto e memória entre diferentes IDEs e CLIs através de um protocolo padronizado (SIP).
Instalação
npx convez-ai frankxai/Starlight-Intelligence-System1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Sincronizar contexto e memória entre Claude, Cursor, Codex e outras ferramentas de IA
- Implementar governança e atestação em sistemas multi-agente
- Construir fluxos de desenvolvimento que mantêm estado consistente entre diferentes ambientes
Veredito

Sistema robusto e bem documentado para coordenação de múltiplos agentes de IA, com protocolo formalizado (SIP) e testes extensivos. Ideal para equipes que precisam de governança real entre ferramentas, mas exige compreensão técnica profunda e setup complexo.
README
A tradução para o português deste README ainda está na fila.
Starlight Intelligence System
The sovereign intelligence substrate — persistent memory, identity, and attested governance for your whole AI fleet.
One brain and one rulebook shared by every coding agent you run: Claude Code, Cursor, Codex, Gemini CLI, OpenCode, Antigravity. Per-tool memory is now table stakes; what your fleet is missing is cross-tool, cross-repo, governed, and attested. That's this repo — 48 agents, 71 auto-activating skills, 6 semantic vaults, an MCP server, and a provenance protocol (SIP). 968 tests keep the claims honest.
60-second start
# 1. Seed the six JSONL vaults (~/.starlight/vaults)
npx -p @arcanea/starlight-intelligence-system starlight init --vaults
// 2. Point any MCP client at them (Claude Code, Cursor, Codex, ...)
{
"mcpServers": {
"starlight": {
"command": "node",
"args": [
"node_modules/@arcanea/starlight-intelligence-system/dist/mcp-server.js",
"--vault-dir", "~/.starlight/vaults"
]
}
}
}
Restart your client: ten sis_* tools in every session — the same vaults, from every CLI you run. Full walkthrough below in Quick start; measured receipts in BENCHMARKS.md.
Two layers, one repo
| Layer | What it is | What lives here | License | Adopt how |
|---|---|---|---|---|
| Substrate (SIP) | A six-layer protocol that lets sovereign parties compose intelligence systems without losing sovereignty. | SIP.md, SIS.md, ALLIANCE.md, STACK.md, VOICES.md, VERTICALS.md, MEMORY.md, REGISTRY.md, SKILL.md, .claude/commands/ | MIT | Read SIP.md, attest with /sip-attest, fork what you need. |
| Operational (reference build) | This repo's working implementation: 48 agents (7-archetype council + specialist tiers + evaluator), 6 semantic vaults, 71 auto-activating skills across 14 domains, MCP server, 14 strategic commands + 100+ slash commands, six-platform adapters. Frank's daily-driver. | agents/, memory/, skills/, commands/, core/, context/, src/ (npm package) | MIT | Install @arcanea/starlight-intelligence-system, run the MCP server, write to your vaults. |
You can adopt just the substrate (fork SIP for your own work), just the operational layer (use the MCP server for AI memory), or the full stack (Frank's reference build, end to end). They are independent.
Operator? Start at SETUP.md — covers
private/instance state, Infisical Path A vs env-var Path B, Windows + Linux cron wiring, cockpit launch, and a smoke test, end-to-end in roughly 30 min.New to the protocol? Don't fork this repo. Fork the SIP adoption kit — eleven markdown files, no code, ship your first attested artifact in 60 seconds. Compose upward when you're ready.
New in v8.2.0 (2026-05-30): First-run experience hardened so the package works for anyone who installs it.
starlight init --vaultsseeds the six JSONL vaults (and the MCP server auto-seeds an empty--vault-diron first boot), so a fresh install is never silently empty.sis_searchis now honestly described as keyword + temporal (no embeddings); a measured retrieval recall@k harness (npm run eval:retrieval) grounds the bm25 claim in CI; and an optionalsqlite-vecsemantic layer is roadmapped indocs/bring-your-own-model.md. SeeCHANGELOG.md § v8.2.0andRELEASING.md.New in v8.1.0 (2026-05-17): Composition Layer primitive declared in
STACK.md— universal IS may compose over its Domain Sub-Stacks via commands at the IS-itself. Wealth IS v0.2 evolved as first composition-layer reference. Crypto Intelligence v0.1 shipped as third reference Domain Sub-Stack (after People + Sound) with Houses-as-sub-systems primitive — House of On-Chain live with 5 commands./blessglobal skill + chronicle infrastructure initialized. 10-IS taxonomy invariant preserved. SeeCHANGELOG.md § v8.1.0+docs/boards/2026-05-17-crypto-investment-spawn.md.
Why It Matters
flowchart TB
Human["Human intent"]
Vaults["6 semantic vaults<br/>JSONL source of truth"]
Index["SQLite + FTS5<br/>rebuildable index"]
MCP["MCP server<br/>10 sis_* tools"]
Adapters["Claude Code · Cursor · Codex · Gemini · OpenCode"]
Repos["Repo-specific agent harnesses"]
Output["Safer, memory-aware agent work"]
Human --> Vaults
Vaults --> Index
Index --> MCP
MCP --> Adapters
Adapters --> Repos
Repos --> Output
Output --> Vaults
| Capability | What It Gives Agents |
|---|---|
| Semantic vaults | Durable memory across sessions and tools |
| Temporal confidence | Old knowledge decays unless reconfirmed |
| Contradiction detection | Conflicting memories become visible |
| Platform adapters | Same substrate across Claude, Cursor, Codex, Gemini, OpenCode |
| MCP server | Tool-native access to memory and retrieval |
| Harness checks | Prompt surfaces stay aligned with reality |
Run the local harness guard:
npm run agents:harness-check
The substrate (SIP)
Starlight Intelligence Protocol — the contract that lets sovereign parties compose intelligence systems without losing sovereignty. Six layers:
- File contract — canonical names and shapes for
SKILL.md,AGENTS.md/VOICES.md,MEMORY.md,CANON.md,SOUL.md,STACK.md,.claude/commands/*, plus.arc/.nea/.skillextensions. - Attestation protocol — verifiable "Built on SIP" attribution via
/sip-attest. Refuses decorative use. - MCP registry standard — how MCP servers declare, compose, version (
mcp.jsonschema). - Command taxonomy — protocol / alliance / vertical / sovereign tiers with explicit decision rights.
- Sovereignty + attribution clause — the non-negotiable social contract.
- Archetype extension — optional canon adoption (Arcanea archetypes available CC-BY-NC).
Canonical URL: starlightintelligence.org/protocol · Source of truth: SIP.md in this repo.
Substrate files at a glance
| File | What it does |
|---|---|
SIP.md | The protocol spec (six layers). |
SIS.md | Substrate map — verticals, composition rules, what SIS does and doesn't provide. |
ALLIANCE.md | The alliance forging method — how 2-5 sovereign nodes compose without collapsing into one entity. |
STACK.md | Recommended sovereign stack (L0-L6). Defaults, not mandates. |
VOICES.md | Five canonical voice archetypes (architect, sovereign-creator, protocol-defender, implementer, overseer). |
VERTICALS.md | Public registry of sovereign verticals + alliance class definitions. |
MEMORY.md | Template for per-instance state. |
REGISTRY.md | MCP server registry. |
SKILL.md | Substrate-layer behavior (what AI adopts when working at this layer). |
.claude/commands/ | 9 reference slash commands (/sip-attest, /alliance-forge, /alliance-reflect, /alliance-decide, /vertical-spawn, /luminor-board, /sovereign-signal, /openclaw-audit, /wealth-dpi). |
Forge an alliance, spawn a vertical
# Pressure-test before committing
/luminor-board "Open-source the agent layer or keep it closed?"
# Forge an alliance under SIP
/alliance-forge trinity "Frank,Ahmad,Logan,Shahvaiz"
# Spawn a vertical IS under SIS
/vertical-spawn anime-legends "anime-aesthetic fiction + character design"
# Attest a shipped artifact
/sip-attest path/to/artifact.md
Every cross-party artifact carries the "Built on SIP" attestation block. Silent composition is a breach.
The operational layer (reference build)
This repo also ships Frank's working implementation — the daily-driver intelligence system that runs on top of SIP. Use it directly, fork it, or replace it entirely with your own layer above the substrate.
Quick start (operational layer, 2 minutes)
Option 1: As an MCP server (recommended for AI tools)
First, seed your vault directory so the system has somewhere to read from
(a fresh install has no ~/.starlight/vaults yet):
npx -p @arcanea/starlight-intelligence-system starlight init --vaults
# → seeds the six JSONL vaults in ~/.starlight/vaults (welcome entry + public examples)
Then point your MCP client at that directory:
{
"mcpServers": {
"starlight": {
"command": "node",
"args": [
"node_modules/@arcanea/starlight-intelligence-system/dist/mcp-server.js",
"--vault-dir",
"~/.starlight/vaults"
]
}
}
}
Restart Claude Code. You now have ten sis_* tools available in every session.
(If you skip the seed step the server still works — it auto-seeds an empty
--vault-dir on first boot so the empty state is never silently broken.)
Option 2: As a library
npm install @arcanea/starlight-intelligence-system
import { StarlightIntelligence } from "@arcanea/starlight-intelligence-system";
import { createAdapter } from "@arcanea/starlight-intelligence-system/adapters";
const sis = new StarlightIntelligence();
sis.initialize();
sis.remember({
content: "Always Read a file before editing — catches stale state",
category: "pattern",
tags: ["workflow", "edit-safety"],
confidence: 0.95,
});
const adapter = createAdapter("claude-code");
const context = await adapter.generate({ vaultDir: "~/.starlight/vaults" });
The six vaults
| Vault | Symbol | Purpose | Example entry |
|---|---|---|---|
| Strategic | ◆ | Business insights, architecture decisions, competitive moats | "Open Core + Founding Circle beats premium tiers at this stage" |
| Technical | ⬡ | Implementation learnings, stack decisions, patterns | "SQLite FTS5 with bm25 beats embeddings for <10k entries" |
| Creative | ✦ | Design preferences, aesthetic rules, voice, lore | "Never Cinzel. Space Grotesk display, Inter body." |
| Operational | ▸ | Workflow patterns, execution lessons, process rules | "Max 2 worktrees. Digest pattern for terminal output." |
| Wisdom | ◎ | Deep principles, truths, cross-domain insights | "Memory that compounds is intelligence that grows" |
| Horizon | ↗ | Vision statements, append-only ledger of human intentions | "Build the substrate that makes AI agents continuous" |
Each vault is a JSONL file. Human-readable. Git-versionable. Greppable.
Where memory lives. The MCP server and
src/retrieval.tsread*.jsonlfiles in your--vault-dir(default~/.starlight/vaults) — that is the one source of truth at runtime. The repo ships two example sets for reference only:public-vault/*.jsonlis the canonical starter content thatstarlight init --vaultscopies into your vault dir;memory/vaults/*.mdare human-readable narrative snapshots and are not read by the engine. Seed once, then your own~/.starlight/vaultsis what counts.
What the operational layer adds on top of SIP
- SQLite hybrid retrieval —
src/retrieval.tsbuilds a rebuildable FTS5 shadow index over JSONL vaults with bm25 ranking. Keyword + temporal today (measured recall in CI vianpm run eval:retrieval); optionalsqlite-vecsemantic layer is roadmapped indocs/bring-your-own-model.md. - Temporal reasoning —
src/temporal.tsadds validity windows and a 90-day confidence half-life. - Contradiction detection —
src/contradiction.tsfinds conflicting entries via word-trigram Jaccard with opposing-signal boosting. - Dreaming —
src/dreaming.tsprocesses session transcripts in the background. - Five platform adapters — Claude Code, Cursor, Codex, Gemini CLI, OpenCode share the same vaults.
- MCP v2 —
src/mcp-server.tsships ten tools over JSON-RPC 2.0 stdio.
MCP tools
| Tool | Description |
|---|---|
sis_vault_search | Free-text search across vaults |
sis_recent_entries | Latest entries from one or all vaults |
sis_stats | Total entry counts per vault |
sis_append_entry | Write a new entry to a vault |
sis_entry_types | List supported vault types and entry categories |
sis_search | Keyword + temporal search (term-overlap score, tag boost, staleness penalty) |
sis_confirm | Touch lastConfirmed on an entry |
sis_invalidate | Mark an entry as expired |
sis_contradict | Flag two entries as contradictory |
sis_stale | List entries not confirmed within a threshold |
Platform adapters
| Platform | Memory file | MCP config | Max tokens |
|---|---|---|---|
| Claude Code | CLAUDE.md | ~/.claude/settings.json → mcpServers | 200,000 |
| Cursor | .cursorrules | Cursor settings → MCP | 128,000 |
| Codex | AGENTS.md | ~/.codex/config.toml | 192,000 |
| Gemini CLI | GEMINI.md | ~/.gemini/settings.json | 1,000,000 |
| OpenCode | AGENTS.md (compact) | ~/.opencode/config.json | 128,000 |
The 7 named agents (operational layer)
The reference build maps SIP's 5 voice archetypes to 7 named runtime agents — Orchestrator, Prime, Architect, Navigator, Sentinel, Weaver, Sage. Full registry: agents/AGENT_REGISTRY.md.
Voice archetypes are abstract; named agents are specific implementations. Anyone forking SIP can choose entirely different agents above the substrate.
Fork patterns
| If you want to... | Take | Leave |
|---|---|---|
| Adopt SIP for your own creator stack | SIP.md, SKILL.md, VOICES.md, .claude/commands/sip-attest.md, .claude/commands/alliance-forge.md | Everything in agents/, memory/, skills/, src/ |
| Use the MCP memory server, no protocol | src/, dist/, npm package | Substrate spec files |
| Forge an alliance under SIP | SIP.md, ALLIANCE.md, VOICES.md, all 9 commands in .claude/commands/ | Operational layer |
| Run the full reference build | All of it | Nothing |
Public canonical URL
starlightintelligence.org/protocol mirrors SIP.md. This repo is the source of truth.
Architecture (operational layer)
┌─────────────────────────────────────────┐
│ JSONL vaults (source of truth) │
│ ~/.starlight/vaults/*.jsonl │
│ human-readable · git-versionable │
└────────────────┬────────────────────────┘
│
│ rebuildable from JSONL
▼
┌─────────────────────────────────────────┐
│ SQLite + FTS5 (shadow index) │
│ bm25 ranking · temporal filters │
└────────────────┬────────────────────────┘
│
│ JSON-RPC 2.0 over stdio
▼
┌─────────────────────────────────────────┐
│ MCP server (10 sis_* tools) │
└────────────────┬────────────────────────┘
│
┌────────────────────┼────────────────────────┐
▼ ▼ ▼ ▼ ▼ ▼
Claude Code Cursor Codex Gemini OpenCode Your tool
Development
git clone https://github.com/frankxai/Starlight-Intelligence-System.git
cd Starlight-Intelligence-System
npm install
npm run build # tsc to dist/
npm test # 82+ orchestrator tests
npm run lint # tsc --noEmit
src/ is under 3,000 lines of TypeScript with zero runtime dependencies outside better-sqlite3. Substrate docs (SIP.md, SIS.md, etc.) are markdown-only — no build dependency.
License
- Code (operational layer + substrate-tier commands): MIT — see
LICENSE. - Substrate spec docs (
SIP.md,SIS.md,ALLIANCE.md,STACK.md,VOICES.md,VERTICALS.md,MEMORY.md,REGISTRY.md,SKILL.md): MIT, sameLICENSE. The "Built on SIP" attestation request is a social-layer convention captured inNOTICE, not a license restriction. - Arcanea canon (if your fork composes with it): CC-BY-NC 4.0, © Arcanea BV. Lives in
frankxai/arcanea-ecosystem; canon is compose-only — not redistributed under this repo's MIT. - Trademarks: ARCANEA, FRANKX, STARLIGHT INTELLIGENCE — registered or in registration, reserved rights.
"Built on SIP" is an attestation phrase, not a trademark. Use of the phrase requires actual SIP composition per /sip-attest rules. Full attribution + canon + trademark summary in NOTICE (Apache-style — LICENSE for legal rights, NOTICE for propagation conventions).
Related
- Arcanea — Creator platform; canon-defining vertical built on SIS
- FrankX — Architect brand; SIP thought leadership
- Public Vaults — Browse and fork vaults
- Protocol page — Canonical SIP spec
- GitHub — Source, issues, discussions
Built on SIP · Starlight Intelligence Protocol · v1.1.1 · v8.2.0 · MIT