The local AI-agent backend

A coding agent or RAG app needs the same handful of local services: a document store, vectors, a cache, a durable job queue, and a scheduler. monlite is all of them in one .db file — no Docker, no Redis, no Qdrant.

Need	monlite
Documents / state	@monlite/core
Semantic memory / RAG	@monlite/vector
Keyword search	@monlite/fts
Cache & locks	@monlite/kv
Durable jobs	@monlite/queue
Scheduling	@monlite/cron

A durable jobs tier

The pattern behind agent task execution: a worker claims a job exactly once (across processes) with compare-and-swap, runs a step, and records progress.

// Claim — cross-process safe; a racing worker gets null, not an error
const job = await jobs.findOneAndUpdate({
  where: { _id: id, status: "pending", version: v },
  data: { $set: { status: "active" }, $inc: { version: 1 } },
  returnDocument: "after",
});

Pair it with @monlite/queue for the work queue, @monlite/kv setNX for single-flight locks/nonces, and compound-unique indexes for idempotency — a durable job/mission/approval engine on a single file, no Mongo or Redis.

RAG in one file

const store = createVectorStore(db);   // @monlite/vector
store.ensureCollection("kb", { dimensions: 384, indexedFields: ["docId"] });
store.upsert("kb", chunks.map((c) => ({ id: c.id, vector: c.embedding, metadata: c.meta })));
store.search("kb", { vector: q, topK: 5, where: { tenantId } }); // scoped, exact

Scoped retrieval (where) is applied inside the KNN, so per-tenant / per-document queries stay exact even over a large corpus. Add @monlite/fts for hybrid keyword+semantic search.

Cross-language pipelines

Because everything is one SQLite file, a Python worker and a Node agent can share it — Python ingests and embeds, Node serves; both hit the same cache and queue. See Python / interop.