Starlightdocs

The context & memory layer

How Starlight combines local sovereign memory (SisGatewayDaemon FTS5 + RRF), build-time Orama indexes, and Streamable HTTP MCP endpoints.

LiveChecked against Starlight-Intelligence-System/AGENTS.md, docs-platform/app/api/mcp/route.ts, docs-platform/app/api/search/route.ts, TRUTH.md
MCP endpoint/api/mcp
1. Verified sourcesBrand MDX + provenance6 SIS memory vaultsGate & craft receiptsskills.graph.json2. Sovereign memorySisGatewayDaemon :5200SQLite FTS5 + vectorsRRF hybrid (Hit@10=1.0)1,106 compiled atoms3. Cloud index & graphOrama build-time indexSupabase pgvector + RLSGraphiti temporal edgeswhen 2+ sources join4. Agent surfaces/api/mcp (HTTP MCP)starlightintelligence.ai/mcp.md & llms-full.txt/api/search per brand

Every page on this platform exposes machine-readable context to coding agents in three ways right now:

  1. Pull feeds: Append .md or send Accept: text/markdown on any URL, or fetch /llms-full.mdx/<brand> for a single-file brand bundle.
  2. Full-text search API: Query /api/search?q=<query>&brand=<brand> against each brand's isolated Orama index.
  3. Streamable HTTP MCP server (/api/mcp): Call list_docs, search_docs, and get_doc over JSON-RPC 2.0 to retrieve Markdown with verified frontmatter citations (checked date and sources).

Two layers: local sovereign memory + cloud docs index

Rather than relying on a single external SaaS database, the estate separates operator memory from public product documentation:

LayerSubstrateVerified performance & scope (2026-10-25)
Local sovereign memory (Starlight-Intelligence-System)Append-only JSONL vaults (Strategic, Technical, Creative, Operational, Wisdom, Horizon) indexed in SQLite FTS5 + vector embeddings via SisGatewayDaemon (127.0.0.1:5200) with Reciprocal Rank Fusion (RRF)1,106 compiled memory atoms, 400 knowledge-graph rows, Eval-50 Hit@10 = 1.000, Recall@5 = 1.000, MRR@10 = 1.000 (p50 = 218.9ms), 0d vault staleness
Public docs & MCP index (docs-platform)Build-time Fumadocs MDX compilation + brand-isolated Orama search + /api/mcp Streamable HTTP serverZero external DB latency; every page carries build-enforced status: verified, checked date, and sources
Cloud relational & vector store (when multi-source)Supabase Postgres + pgvector + Row Level Security (RLS)Used when joining documentation chunks with dynamic cloud records (DemandSignal, gate receipts)
Temporal knowledge graph (when edge queries earn it)Graphiti on FalkorDB (bi-temporal validity windows)Reserved for cross-artifact lineage ("which receipt superseded which release")
See the MCP tools reference

Why local FTS5 + RRF and build-time Orama come before external SaaS

  1. Zero drift from Git: Documentation lives in version-controlled MDX with a Zod provenance gate in lib/source.ts. If a page's status: verified frontmatter lacks a checked date or sources list, next build fails immediately.
  2. Progressive disclosure over heavy RAG: As recorded in TRUTH.md §3b (arXiv 2607.17598), single-level index-to-passage lookup (llms.txt → .md page or search_docs → get_doc) outperforms multi-hop vector routing at corpus scales under 2M words while preserving exact file citations.
  3. Sovereign privacy boundary: Private operator notes stay in starlight-private-memory and local SisGatewayDaemon loopback (127.0.0.1:5200 with privacy tag scrubbing), never leaking into public cloud indexes.

Use Starlight from your own agent code

Connect an agent built with the Vercel AI SDK, the Claude Agent SDK, or Vercel Eve to the Starlight MCP endpoints.

After Smithery - move your MCP servers

Smithery's connection services end on 31 October 2026. Swap each Smithery address for the vendor's own, or host your server yourself on Cloudflare Workers and list it in the official MCP Registry.

On this page

Two layers: local sovereign memory + cloud docs indexWhy local FTS5 + RRF and build-time Orama come before external SaaS