nlqdb
Analytical memory for AI agents: a real Postgres your agent connects to over MCP and queries in plain English — GROUP BY, JOIN, aggregate over what it remembered, not just top-k vector recall.
README
nlqdb — analytical memory for AI agents
Memory your agent can query, not just recall — a real database it reaches over MCP.
Connect nlqdb to Claude, Cursor, Codex, or any MCP host. Your agent writes
typed rows as it learns, then asks questions in plain English — GROUP BY,
JOIN, aggregate over what it remembered. A vector store returns the top-k
similar chunks; nlqdb runs the query a similarity index structurally can't.
The LLM never emits SQL: it returns a typed plan, our compiler emits the
parameterised statement, and you see the exact SQL every time.
It's also a natural-language database for any app. You write HTML; each component asks for what it wants in plain English; nlqdb infers the schema, writes the SQL, runs it, and renders the result. There is no backend for you to build.
Two actions. That's the whole product:
- Create a database — one word: a name (or a goal).
- Talk to it in plain English.
<script src="https://elements.nlqdb.com/v1.js" type="module"></script>
<nlq-data
goal="the 5 newest orders, with customer and item"
api-key="pk_live_xxx"
template="table"
refresh="10s"
></nlq-data>
That's the entire backend for a live order list — no API to write, no schema to define, no JSON to parse. Engine choice (Postgres / Mongo / Redis / DuckDB / pgvector / …), schema inference, indexing, backups, and auto-migration between engines based on your real workload are background concerns you never have to see.
Status — early, open
nlqdb is early and built in the open, but fully public — no gate, no
invite code. The marketing site, the /v1/ask pipeline, the <nlq-data> /
<nlq-action> elements, the chat app, the TypeScript SDK, the hosted MCP
server, and the nlq CLI are all live in some form (see the surface table
below). Natural-language → SQL accuracy is still climbing toward our public
bar (BIRD ≥ 0.65, Spider 2.0 ≥ 0.75 on the free model chain), so answers can
be wrong — every response carries a confidence signal and the SQL it ran.
Use it
Connecting an agent over MCP? On Claude Code, one marketplace add wires the hosted server and both memory skills in a single step:
/plugin marketplace add nlqdb/nlqdb
/plugin install nlqdb-memory@nlqdb
On any other MCP host, give your agent memory
with one browser-OAuth approval; headless hosts skip the browser with
npx -y @nlqdb/mcp (0.1.1) and an sk_mcp_* MCP key
(MCP setup). @nlqdb/sdk (0.2.2) and
@nlqdb/mcp (0.1.1) are both published and importable from npm.
The 60-second walkthrough — plain HTML, CLI, and ten framework wrappers —
lives at docs.nlqdb.com. Start with the
HTML tutorial or the
CLI tutorial.
You don't generate an API key separately: describe your database at
nlqdb.com, and the chat hands you a
<nlq-data> snippet with the key already inlined.
Examples
examples/ — minimal scaffolds in plain HTML, Next.js,
Nuxt, SvelteKit, Astro, plus a CLI-only walkthrough. Each is the smallest
valid integration around one <nlq-data> element or one CLI session.
What makes it different
Four things every release has to move, none allowed to regress
(GLOBAL-025):
- Engine quality — natural-language → SQL accuracy (measured continuously on BIRD + Spider 2.0 + an internal eval), plus the multi-engine layer that moves your data to the right engine for your workload.
- Onboarding — landing to first answer in under a minute, no card, no config.
- UX — see the diff before any write, see the SQL behind every answer, refuse rather than guess when confidence is low.
- Performance — sub-400 ms cached, sub-1.5 s cold.
The bet: get this right on free, open models and it only gets better on frontier ones — the scaffolding compounds with whatever model is underneath.
Models & plans
- Free forever on the built-in open-model chain — queries, embeds, and the elements, no card required.
- Bring your own LLM key (Anthropic / OpenAI / Gemini / Grok / OpenRouter) on any tier, at no markup.
- Hosted premium models on paid plans, when you'd rather not manage a key of your own.
- Self-host the source — the engine, CLI, MCP server, and SDKs are source-available under FSL-1.1-ALv2: free to self-host for any non-competing use, bring your own LLM key, no per-call fees. The license auto-converts to Apache 2.0 two years after each release.
Paid plans aren't live yet. The full model strategy is in
GLOBAL-026.
Surfaces at a glance
| Surface | Status | Where |
|---|---|---|
HTTP API (POST /v1/ask, POST /v1/run) |
✓ shipped | apps/api/src/ask/** |
<nlq-data> + <nlq-action> elements |
✓ shipped (v0.1) | packages/elements/** |
@nlqdb/sdk (TypeScript) |
✓ shipped (incl. runSql) — installable from npm (0.2.2) |
packages/sdk/** |
| Framework wrappers (React / Next / Vue / Nuxt / Svelte / SvelteKit / Astro / Solid + Swift) | ~ built + CI-tested; npm / SPM publish pending | packages/{react,next,…}/** |
Chat app nlqdb.com/app |
✓ shipped | apps/web/** |
Hosted MCP server mcp.nlqdb.com/mcp |
✓ shipped (host auto-detect pending) | apps/mcp/**, packages/mcp/** |
Local stdio MCP server @nlqdb/mcp |
✓ shipped (0.1.1) — npx -y @nlqdb/mcp with an sk_mcp_* key |
packages/mcp/** |
| Droppable agent-memory artifacts (AGENTS.md · Claude Code skill + plugin · Cursor rules · Codex config) | ✓ shipped — /plugin marketplace add nlqdb/nlqdb installs the server + skills in one step |
apps/web/public/agent-artifacts/** |
nlq CLI (Go) |
✓ shipped (core verbs; device-login pending) | cli/** |
Full integration matrix in docs/progress.md.
Roadmap
The detailed plan below is a summary; the canonical phase plan and exit
gates live in docs/phase-plan.md. Legend:
✓ shipped · ~ in progress · ◯ planned.
Phase 0 — Foundations ✓
Worker skeleton · KV + D1 + R2 bindings · Neon adapter + OTel · LLM router
(free chain) · Better Auth (GitHub + Google + magic link) · /v1/ask
end-to-end · events queue + drain · Stripe webhook · CI/CD + PR preview
environments.
Phase 1 — On-ramp
A stranger lands on nlqdb.com, creates a DB in plain English, embeds it,
and shares the link — in under 60 seconds, no card, no config.
- ✓ Marketing site (Astro, live at
nlqdb.com) - ✓
<nlq-data>+<nlq-action>elements (v0.1) - ✓ Sign-in — magic link + GitHub + Google
- ✓ Chat surface — streaming three-part response (answer / data / trace), anonymous mode
- ✓ Anonymous mode — 72h token, adopted onto your account on sign-in
- ✓ Hosted db.create pipeline (table-card embeddings stubbed pending the pgvector slice)
- ✓ API keys dashboard (
/app/keys) - ◯ Hello-world tutorial polish
Phase 1.5 — Trust + telemetry
- ✓ Diff preview on writes + visible SQL trace on every response
- ✓ Demand-signal telemetry on every "not yet" path
- ◯ Confidence floor (refuse-on-low-confidence) — lands with quality-eval
Phase 2 — Distribution (agent + developer surfaces)
- ✓ Hosted MCP server (
mcp.nlqdb.com/mcp) — host auto-detect pending; local stdio@nlqdb/mcp@0.1.1is on npm, sonpx -y @nlqdb/mcpwith ansk_mcp_*key is a headless route in with no browser consent step (/agentsnow carries it; the per-host install panel is still OAuth-only). On Claude Code,/plugin marketplace add nlqdb/nlqdbinstalls the server + both memory skills in one step - ✓ CLI
nlq(Go) — core verbs + raw-SQL escape hatch; device-login + chat REPL pending - ✓
@nlqdb/sdk— basic methods +runSql; published and importable from the registry (0.2.2) - ~ Framework wrappers + native Swift package — built + CI-tested; npm / SPM publish pending
- ✓ Quality-eval harness (BIRD + Spider 2.0, manual on-demand) — the free-vs-frontier accuracy delta is the headline KPI
- ~ Bring-your-own-LLM dispatch — HTTP lane live; remaining surfaces in progress
- ◯ CSV upload in chat
- ~ Docs-site reference completeness — SDK + framework-wrapper guides,
an enumerable error-code reference, and a build-time
/llms.txtfor agents now live; tutorial polish remains - ◯ Custom domains for embeds
Phase 3 — Multi-engine engine (the moat)
- ◯ Workload analyzer → migration orchestrator
- ◯ ClickHouse / DuckDB / Redis as additional engines
- ◯ Dual-read verification
- ◯ Hosted-premium model lane (demand-gated)
Phase 4 — Beyond v1
- ~ Bring-your-own Postgres / ClickHouse — connect path live end-to-end
(
POST /v1/db/connect+ web UI, CLI, SDK, query dispatch); prod-gated on theBYO_SECRET_KEKsecret - ◯ SSO (SAML / OIDC), audit-log export, per-org quotas
- ◯ EU data residency, VPC peering, SOC 2
Develop locally
git clone git@github.com:nlqdb/nlqdb.git && cd nlqdb
scripts/bootstrap-dev.sh # installs everything, pulls Ollama models, seeds .envrc
scripts/login-cloud.sh # signs you into cloud providers that have a CLI flow
bootstrap-dev.sh stands up the whole toolchain in one shot — Bun, Node
20+, Go 1.25+, uv; Biome / gofumpt / golangci-lint / ruff; lefthook git
hooks; the cloud CLIs (wrangler, flyctl, stripe, gh); a local Ollama so the
LLM router works offline; and a .envrc with self-generated dev secrets.
Details in
docs/history/infrastructure-setup.md §8.
Day-to-day:
bun run fix # biome format + lint --write (most issues)
bun run check:all # biome + golangci-lint + ruff (what CI runs)
bun run hooks:run # run pre-commit hooks against staged files
End-to-end tests (manual trigger)
E2E coverage is persona-driven and manually triggered so cost stays
inside the free-tier envelope — one workflow_dispatch workflow per
surface:
gh workflow run e2e-opencheck.yml # web — live LLM, Neon branch, Workers preview
gh workflow run e2e-cli.yml # Go testscript, hermetic
gh workflow run e2e-sdk.yml # vitest + cassettes, hermetic
gh workflow run e2e-mcp.yml # InMemoryTransport protocol tests, hermetic
gh workflow run e2e-examples.yml # Playwright across HTML/Next/Astro/Nuxt/SvelteKit
gh workflow run e2e-examples.yml -f live=true # + staging for the curl + CLI shell smokes
Run the hermetic surfaces locally without GitHub:
( cd tests/e2e/cli && go test ./... )
( cd tests/e2e/sdk && bun install && bun run test )
( cd tests/e2e/mcp && bun install && bun run test )
( cd tests/e2e/examples && bun install && bun run install:browsers && bun run test )
Only execution is manual: tests/e2e/{sdk,mcp,examples} live outside the root
workspace, so CI's typecheck-e2e job tscs them on every PR — the free
backstop against a suite that compiles today and rots before the next dispatch.
Conventions, persona mapping, and cassette governance are in
docs/features/e2e-coverage/FEATURE.md.
Docs & reference
docs/architecture.md— system design (auth, pricing, the $0 stack, model selection, hosted db.create, hello-world).docs/phase-plan.md— canonical phase plan and exit gates.docs/decisions.md— cross-cuttingGLOBAL-NNNdecisions; per-feature records live underdocs/features/.docs/performance.md— SLOs, latency budgets, span/metric catalog.docs/competitors.md— competitive landscape.
Community & legal
- CONTRIBUTING.md — dev setup, branch naming, commits, CLA flow.
- CODE_OF_CONDUCT.md — Contributor Covenant 2.1. Reports to
conduct@nlqdb.com. - SECURITY.md — vulnerability disclosure (
security@nlqdb.com). 90-day fix target. - SUPPORT.md — where to ask questions and what we don't (yet) offer.
- CLA.md — Contributor License Agreement, signed once via the bot on your first PR.
- TRADEMARKS.md — what you can and can't do with the nlqdb name and logo.
- SUBPROCESSORS.md — third-party services that may process personal data on our behalf.
- IMPRESSUM.md — Swiss UWG-mandated operator disclosures.
- Privacy policy and terms of service: nlqdb.com/privacy · nlqdb.com/terms.
License
FSL-1.1-ALv2 — Functional Source License, Apache 2.0 future license. Source-available for any non-competing use; auto-converts to Apache 2.0 two years after each release. (Pattern used by Sentry, Convex, and others.)
nlqdb™ is an unregistered trademark of the project's licensor. See
TRADEMARKS.md for usage guidelines.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.