Starlight Intelligence System MCP server
Commit Nov 14, 2026Provides a shared semantic memory substrate and orchestration layer for AI agents to maintain persistent, attested knowledge across different tools.
Also known as Starlight-Intelligence/Starlight-Intelligence-System
- Stars
- 10
- Tools
- 13
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 14, 2026
- License
- MIT
What is Starlight Intelligence System MCP?
Starlight Intelligence System provides a shared semantic memory substrate and orchestration layer for AI agents to maintain persistent, attested knowledge across different tools. It exposes 13 tools, such as sis_vault_search, sis_recent_entries, sis_stats.
Summary generated from the project's README
About Starlight Intelligence System MCP server
Starlight Intelligence System provides a shared semantic memory substrate designed to unify AI agents across different environments. It allows tools like Claude Code, Cursor, Codex, Gemini CLI, and Antigravity to read from and write to the same six semantic vaults. This ensures that knowledge remains persistent and attested rather than being lost between different sessions or tools.
The system is implemented as a local MCP server that connects to your existing AI clients. Users can deploy it by cloning the repository and pointing their MCP configuration to the server file, specifying a local vault directory. It functions as a sovereign runtime where private memory and orchestration remain under your control.
Beyond simple memory, the system includes an orchestration layer and a reference implementation known as SIS. This includes thirteen specific sis_* tools that appear in your client once connected. It is built upon the Starlight Intelligence Protocol, which defines how memory, identity, and governance are handled across an entire fleet of agents.
What you can do with it
- 1Synchronize memory across toolsProvide a single source of truth so that different AI agents share the same context and decisions.
- 2Maintain persistent agent knowledgeEnsure that information and decisions made in one session are available to future agent interactions.
- 3Deploy a local memory serverSet up a private MCP server to manage semantic vaults on your own machine.
- 4Automate agent skill creationUse the Foundry to turn specific intents into typed task envelopes and compiled skills.
Written from the project's README. Every use case is checked against a quote from it.
Starlight Intelligence System MCP tools 13
Detected from the READMEsis_vault_searchsis_recent_entriessis_statssis_append_entrysis_entry_typessis_searchsis_confirmsis_invalidatesis_contradictsis_stalesis_goal_statussis_goal_update
| sis_vault_search | — | |
| sis_recent_entries | — | |
| sis_stats | — | |
| sis_append_entry | — | |
| sis_entry_types | — | |
| sis_search | — | |
| sis_confirm | — | |
| sis_invalidate | — | |
| sis_contradict | — | |
| sis_stale | — | |
| sis_goal_status | — | |
| sis_goal_update | — |
Show all 13 toolsShow fewer tools
sis_goal_log
| sis_goal_log | — |
Automated checks
Ran Nov 14, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablestarlight-intelligence/starlight-intelligence-system is public and reachablegithub.com/starlight-intelligence/starlight-intelligence-sys ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-13github.com/starlight-intelligence/starlight-intelligence-sys ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model13 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNo tool named like delete, drop or execPass
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubStarlight Intelligence System
One shared brain for your entire AI fleet — across every tool and every repository.
Persistent memory. Attested knowledge. Sovereign control.
Claude Code, Cursor, Codex, Gemini CLI, OpenCode, Antigravity — they all read from and write to the same six semantic vaults, with full provenance. No more explaining the same thing to every agent. No drift. No lost decisions.
Starlight Intelligence is an AI systems project led by Frank Riemer in Amsterdam. This repository contains its public protocol and reference source. About the project and product · Released open-source foundations · Contribute and explore the community proposal.
Choose your first result
Starlight is modular. Use the smallest part that solves your problem; a hosted account is not required to explore the protocol or run the local reference implementation.
| Your goal | Start here | Scope |
|---|---|---|
| Understand the open contract | Read SIP | Protocol and interoperability; no installation |
| Share memory across AI tools | 60-second local start | Local vaults and MCP server; bring your own tools |
| Explore the public interface | Deploy Starlight Explorer | Read-only Explorer, not the private agent runtime |
| Improve the platform | Contribution guide | Tests, fixes, adapters, documentation and protocol proposals |
Canonical source today: this frankxai/Starlight-Intelligence-System repository owns issues, pull requests and releases. Same-named copies in other organizations are not alternate contribution targets. If repository ownership changes, follow the verified canonical destination announced here; do not open duplicate PRs against mirrors.
Deploy Starlight Explorer
This creates your own Git-connected copy of the public Starlight Explorer: the protocol, architecture, research, knowledge maps, and read-only public vault surfaces. It builds the site/ Next.js application with zero required secrets.
Private memory, the MCP server, agent workers, credentials, and orchestration remain in your local sovereign runtime. See the exact deployment boundary or run the complete system.

Claude Code, Cursor, Codex, Gemini, and OpenCode — all reading the exact same attested memory from one substrate.
Starlight is two things in one repository:
- SIP — the open Starlight Intelligence Protocol: a clean contract for sovereign memory, identity, and governance that anyone can adopt.
- SIS — the reference implementation: source-counted agent definitions, auto-activating skills, six semantic vaults, a production MCP server, and the repeatable Estate Factory for running full sovereign agent armies.
The current 151 agents and all other fast-moving counts are dated in metrics/current.json and enforced by both the agent harness and site build.\n\nUse just the memory layer in your current tools. Adopt the protocol. Or run the complete system Frank uses every day.
Purpose
Starlight Intelligence System (SIS) is the sovereign AI substrate for memory, orchestration, skills, governance, and evals across Claude Code, Codex, Cursor, Gemini, and Antigravity. Dual-layer: SIP substrate you can adopt/fork, plus this repo's reference operational build (vaults, MCP, agents).
How People Experience It
- Builders: 60-second start → forge alliances → spawn vertical intelligence systems → attest artifacts.
- Operators: six vaults under
~/.starlight/vaults, MCP tools, platform adapters, named agents. - Public narrative: constellation / ecosystem maps and Health Intelligence (and sibling verticals) downloads.
- Humans land on: README → 60-second start → substrate files → operational quick start.
How Agents Explore It
- Read
CLAUDE.md(deepest) andAGENTS.md(SIP file-contract / Codex entry). - Check
agents/AGENT_REGISTRY.mdbefore inventing agents. - Memory first: vaults exist — query, then write durable notes for future sessions.
- Run harness gates before claiming production readiness; prefer proof over console language (#35).
- Cross-links: starlight-memory · starlight-evals · starlight-swarm · ACOS · FrankX.
Foundry: portable capabilities for ChatGPT and Codex
Starlight Foundry turns intent into a typed Task Envelope, resolves existing capabilities, compiles the smallest justified skill or agent system, and issues an evidence receipt before promotion.
/forge skill <brief>
/forge agent <brief>
/forge swarm <brief>
/forge vertical <brief>
/forge plugin <brief>
/prove <package-or-artifact>
/evolve <package-or-receipt>
The skills-only starlight-foundry plugin works as the portable ChatGPT Work/Codex surface. Agent creation fails unless a durable decision-right, memory, tool, ownership, or ongoing-trigger boundary is proven.
Read the Foundry architecture · Inspect the contracts · Run the demo
Starlight Agent Designer
The skills-only Agent Designer plugin turns a human job into a value case, architecture, scoped contract, activation prompt and evaluation plan. Its constitutional foundations include human agency, truthfulness, craftsmanship, privacy, stewardship, creativity and scientific rigor. A tested in-process gateway reference demonstrates bounded preparation; live runtime enforcement remains a separate integration. Read the design and activation guide.
Usefulness
- Humans: portable sovereign memory + multi-agent fleet that compounds.
- Agents: shared cognitive architecture, SAGE self-healing patterns, alliance/vertical forge.
- Portfolio role: substrate under GenCreator/ACOS/Arcanea ops; mesh and Wealth OS drafts stay issue-tracked until harness-proven.
60-second start
git clone https://github.com/frankxai/Starlight-Intelligence-System
cd Starlight-Intelligence-System && npm install && npm run build
node dist/cli.js init --vaults
npm serves
@arcanea/starlight-intelligence-system@8.5.1as checked on 10 October 2026. This checkout uses Node 22 or 24. The modular packages are release candidates; the operational SDK keeps its own contracts and sanitizer until a separately versioned root migration is published.
Then add to your MCP config (Claude Code, Cursor, Codex, Gemini, etc.), pointing at the dist/mcp-server.js inside your clone:
{
"mcpServers": {
"starlight": {
"command": "node",
"args": ["/path/to/Starlight-Intelligence-System/dist/mcp-server.js", "--vault-dir", "~/.starlight/vaults"]
}
}
}
(If you installed via npm instead, the path is node_modules/@arcanea/starlight-intelligence-system/dist/mcp-server.js — but see the version note above.)
Restart your client. Every agent now shares the same memory. Thirteen sis_* tools appear instantly (verified by JSON-RPC handshake, 2026-07-12).
Two layers, one repo
| Layer | What it is | Best for |
|---|---|---|
| SIP (Protocol) | Open standard. Six clean layers for memory, attestation, commands, and sovereignty. | Forking the protocol or building your own systems |
| SIS (Reference) | Complete working system: source-counted agents and skills, six vaults, MCP server, Estate Factory. | Running it as your daily multi-agent brain |
Use the protocol alone, the memory layer alone, or the full reference implementation. They are designed to compose.
Getting started locally — SETUP.md (30-minute full setup including private state). Just want the protocol spec — use the lightweight SIP kit instead.
See CHANGELOG.md for history.
Why It Matters
flowchart TB
Human["Human intent"] --> Vaults["6 semantic vaults (JSONL truth)"]
Vaults --> Index["SQLite + FTS5 index"]
Index --> MCP["MCP server (13 sis_* tools)"]
MCP --> Adapters["Every CLI you use"]
Adapters --> Work["Safer, consistent, memory-aware work"]
Work --> Vaults
The difference:
- One source of truth instead of N fragmented contexts
- Knowledge that compounds instead of resetting every session
- Attestation so you always know what came from where
- The same memory works in Claude, Cursor, Codex, Gemini — simultaneously
We made the hard parts (temporal decay, contradiction detection, cross-tool sync) automatic.
For production sovereign systems
Starlight includes a complete Estate Factory pattern for running real "Mind + Mesh + Steward" agent armies at scale.
- Mind: Your grounded intelligence layers
- Mesh: Coordinated swarms with routing, synthesis, and operations
- Steward: Health monitoring, evolution, and long-term value
A 4-layer blueprint plus reusable templates lets you stand up new estates quickly while keeping them 100% yours.


Full workflow and templates: docs/delivery/estate-army-commissioning-workflow.md.
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 substrate-tier slash commands (/sip-attest, /alliance-forge, /alliance-reflect, /alliance-decide, /vertical-spawn, /luminor-board, /sovereign-signal, /openclaw-audit, /wealth-dpi) among 121 total command files. |


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 "architect,sovereign-creator,protocol-defender,implementer"
# 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):
# From a git clone of this repo:
node dist/cli.js 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": [
"/path/to/Starlight-Intelligence-System/dist/mcp-server.js",
"--vault-dir",
"~/.starlight/vaults"
]
}
}
}
For an npm installation, use node_modules/@arcanea/starlight-intelligence-system/dist/mcp-server.js instead; the clone path above is required for the current v8.x workflow.
Restart Claude Code. You now have thirteen 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@8.5.1
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. - Six platform adapters — Claude Code, Cursor, Codex, Gemini CLI, OpenCode, Antigravity share the same vaults.
- MCP v2 —
src/mcp-server.tsships thirteen tools over JSON-RPC 2.0 stdio. - Multi-harness runtime bridge & context continuity —
@arcanea/starlight-intelligence-system/runtime-bridgeandsrc/context-bridge.tsprovide bounded context capture, atomic journal replacement, revision-scoped handoffs, cooperative PID/token locks, and typed review queues. - Jev router policy primitive —
src/runtime-bridge/jev-router.tsprepares public-only requests from a host-supplied classification and model pool. Post-response checks cannot undo transmission or certify cost.
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 |
sis_goal_status |
Read the current state of a tracked goal |
sis_goal_update |
Update a tracked goal with attested state |
sis_goal_log |
Append progress evidence to a tracked goal |
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) |
~/.config/opencode/opencode.json |
128,000 |
| Antigravity | AGENTS.md |
Built-in native tools & MCP | 1,000,000+ |
| Grok | AGENTS.md / GROK.md |
grok-harness-adapter |
131,000 |
Multi-harness runtime bridge & context continuity
The @arcanea/starlight-intelligence-system/runtime-bridge export provides ContextBridge and ReviewQueue for bounded context capture and revision-bound handoffs. The host owns scheduling, provider admission and verification of the receiving checkout.
- Journal writes retain prior events and replace the JSONL file atomically under an owned lock. UUID scratch-file cleanup runs under that same lock; constructing a reader cannot delete a pending writer's file.
- Handoffs select the latest checkpoint for the requested repository and revision. Captured records remain untrusted data; they do not authenticate a repository or establish the receiving checkout's state.
- Cooperative local locks retain a random ownership token, host and PID. Recovery requires the inspected token and an exited local owner. Live, remote and unverifiable owners remain protected; age is never takeover authority.
- Review requests bind the repository, revision, policy and distinct maker/checker providers. Receipts record reported commands, findings and limitations. They are not cryptographically signed or independent proof that those commands ran.
- Jev requests require a host-approved public classification and current model pool. Returned metadata is an untrusted provider report; policy preparation alone does not prove data classification, routing or billed cost.
The source recovery changes require a subsequent root release. Published root 8.5.1 is immutable; merging source does not update its registry archive.
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, the 9 substrate 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.
Starlight Intelligence System MCP questions
What tools does Starlight Intelligence System MCP provide?
It exposes 13 tools, including sis_vault_search, sis_recent_entries, sis_stats, sis_append_entry, sis_entry_types.
What tools can use Starlight memory?
The system is designed to work across Claude Code, Cursor, Codex, Gemini CLI, OpenCode, and Antigravity.
How do I add Starlight to my MCP client?
You can add it to your MCP configuration by pointing to the dist/mcp-server.js file in your local clone using a node command.
Does Starlight require a hosted account to use?
No, a hosted account is not required to explore the protocol or run the local reference implementation.
What is included in the SIS reference implementation?
It includes source-counted agent definitions, auto-activating skills, six semantic vaults, a production MCP server, and the Estate Factory.