# Duxxdb > The database built for AI agents - hybrid vector + BM25 + structured, durable, with RESP / MCP / gRPC servers. Apache 2.0. - **Type:** MCP server - **Install:** `agentstack add mcp-bankyresearch-duxxdb` - **Verified:** Pending review - **Seller:** [bankyresearch](https://agentstack.voostack.com/s/bankyresearch) - **Installs:** 0 - **Category:** [Databases](https://agentstack.voostack.com/c/databases) - **Latest version:** 0.1.0 - **License:** Apache-2.0 - **Upstream author:** [bankyresearch](https://github.com/bankyresearch) - **Source:** https://github.com/bankyresearch/duxxdb - **Website:** https://www.duxx.ai/duxxdb ## Install ```sh agentstack add mcp-bankyresearch-duxxdb ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About Hybrid retrieval (vector + BM25 + structured), agent-native primitives (`MEMORY` / `TOOL_CACHE` / `SESSION`), and embedded-or-server deployment from one Rust codebase. Speaks RESP2/3, gRPC, and MCP out of the box. [](https://github.com/bankyresearch/duxxdb/actions/workflows/ci.yml) [](LICENSE) [](https://www.rust-lang.org/) [](https://github.com/bankyresearch/duxxdb/actions/workflows/ci.yml) [](docs/PROJECT_OVERVIEW.md) [](docs/) [**Install**](#install) · [**Quickstart**](#quickstart) · [**Why**](#why-duxxdb) · [**Features**](#features) · [**Integrate**](docs/INTEGRATION_GUIDE.md) · [**FAQ**](docs/FAQ.md) · [**Architecture**](#architecture) · [**Benchmarks**](#benchmarks) · [**Roadmap**](docs/ROADMAP.md) · [**Contribute**](CONTRIBUTING.md) --- ## TL;DR ```bash docker run -d --name duxxdb -p 6379:6379 ghcr.io/bankyresearch/duxxdb:latest redis-cli -p 6379 REMEMBER alice "I lost my wallet at the cafe" redis-cli -p 6379 RECALL alice "wallet" 3 # 1) 1) (integer) 1 # 2) "0.032787" # 3) "I lost my wallet at the cafe" ``` DuxxDB just spoke RESP — `redis-cli`, `valkey-cli`, `redis-rs`, `node-redis`, `go-redis`, all work unchanged. Now switch to the [Quickstart](#quickstart) for Python, gRPC, or MCP. > **Status:** self-hosted beta — feature-complete through Phase 7 > with auth + RBAC, native TLS, health, Prometheus, graceful shutdown, > persistence, mTLS client verification, configurable protocol limits, > offline snapshot/restore, importance-based eviction, and deletion-safe > recall (graph compaction). See the [Roadmap](docs/ROADMAP.md). > > Multi-tenancy, the control plane, replication/HA, and the Cloud Console > live in **DuxxDB Cloud** (a separate managed-services product); this repo > is the Apache-2.0 single-node engine. See [LICENSING.md](LICENSING.md). --- ## The Duxx Stack DuxxDB is the **storage engine** half. The **framework** half is [**duxx-ai**](https://github.com/bankyresearch/duxx-ai) — an Apache 2.0 Python SDK for building, fine-tuning, orchestrating, and governing AI agents: ``` ┌─ duxx-ai (Python) ────────────────────────────────────┐ │ agents · graph · crew · tools · memory · guardrails │ │ RBAC · observability · fine-tune · adaptive routing │ └────────────────────┬──────────────────────────────────┘ │ uses (memory / traces / prompts / eval / cost) ┌────────────────────▼──────────────────────────────────┐ │ DuxxDB (Rust) — storage + retrieval + observability │ │ hybrid recall · TLS · MCP · gRPC · RESP · Parquet │ └───────────────────────────────────────────────────────┘ ``` | You're building... | Reach for | |---|---| | The agent itself (Python) | [`duxx-ai`](https://pypi.org/project/duxx-ai/) — agents, crews, tools, governance | | The storage / retrieval layer | **`duxxdb`** (this repo) — single binary, six client surfaces | | Both together | `pip install duxx-ai duxxdb` — the full Apache 2.0 agent stack | **Why one stack, two projects?** Rust gets you sub-ms hybrid recall + TLS + multi-platform binaries. Python gets you the agent-developer ergonomics — typed primitives, LLM SDKs, tools, governance — without dragging C++ deployment headaches into the storage tier. Each side plays to its strengths; they ship independently with semver. --- ## Why DuxxDB? A typical agent stack today glues three databases together: ``` ┌──────────┐ ┌──────────┐ ┌──────────────┐ │ Redis │ │ Qdrant │ │ Postgres │ │ session │ │ vectors │ │ facts/users │ │ ~0.5ms │ │ ~10ms │ │ ~1-5ms │ └────┬─────┘ └────┬─────┘ └──────┬───────┘ ▼ ▼ ▼ 3 round-trips per turn 3 client libs · 3 sets of credentials eventual consistency between them ``` The tax shows up in three places: | Tax | Cost | |---|---| | **Latency** | 3 hops = 300 µs–3 ms before any work. Voice bots blow their 200 ms budget. | | **Consistency** | A user message, its embedding, and its session metadata land in three stores at three times. | | **Glue code** | Cross-store transactions, retries, drift — the silent killer of agent reliability. | **DuxxDB collapses the three into one engine** with first-class agent primitives, a single hybrid query plan (vector ANN + BM25 + structured filters fused via Reciprocal Rank Fusion), and six integration surfaces out of one codebase. Pure Rust, no GC pauses. --- ## Features #### Hybrid retrieval - HNSW vector ANN ([`hnsw_rs`](https://crates.io/crates/hnsw_rs)) - BM25 full-text ([`tantivy`](https://github.com/quickwit-oss/tantivy)) - Structured filters - **All three fused in one query plan** (RRF, k=60) #### Agent-native primitives - `MEMORY` — sticky long-term recall with **importance decay** - `TOOL_CACHE` — exact + **semantic-near-hit** lookup (cosine ≥ 0.95) - `SESSION` — sliding-TTL KV with lazy eviction - All first-class types — not generic tables #### Six integration surfaces - **Embedded** Rust crate - **Python** wheel (PyO3, abi3-py38, one wheel for 3.8–3.13) - **Node / TypeScript** native module (napi-rs v2) - **RESP2/3** TCP server (Valkey/Redis-compatible) - **gRPC** with streaming `Subscribe` (tonic) - **MCP stdio** for Claude / GPT / Cline agents #### Production hardening (Phase 6.1 + 6.2) - Token auth (`--token` / `DUXX_TOKEN`) - Role-based RESP API keys (`DUXX_AUTH_KEYS`) with read/write/admin roles - JSON-lines security audit log (`DUXX_AUDIT_LOG`) - **Native TLS** on RESP + gRPC (rustls) - **mTLS client verification** with `--tls-client-ca` - **Configurable RESP resource limits** (`DUXX_MAX_*`) - Connection and per-connection command-rate caps - Prometheus `/metrics` + `/health` - gRPC `grpc.health.v1.Health` protocol - Graceful shutdown drain - **Memory cap + importance-based eviction** - Offline `duxx-snapshot create/verify/restore` - Apache Parquet cold-tier export --- ## Install Postgres-style — pick your platform. | Platform | Command | Details | |---|---|---| | 🐳 **Docker** (any OS) | `docker run -p 6379:6379 ghcr.io/bankyresearch/duxxdb:latest` | [Docker guide](docs/INSTALLATION.md#docker-single-command) | | 🐳 **Docker Compose** (production) | `cd packaging/docker && cp .env.example .env && docker compose up -d` | [Compose guide](docs/INSTALLATION.md#docker-compose-production-grade) | | 🐧 **Debian / Ubuntu** | `sudo apt install ./duxxdb_*.deb` | [.deb install](docs/INSTALLATION.md#linux--debian--ubuntu-apt-deb) | | 🐧 **RHEL / Fedora / Rocky** | `curl -fsSL …/install.sh \| sudo sh` + systemd unit | [.rpm install](docs/INSTALLATION.md#linux--rhel--fedora--rocky--alma-dnf-rpm) | | 🍏 **macOS** | `brew install bankyresearch/duxxdb/duxxdb` | [Homebrew](docs/INSTALLATION.md#macos--homebrew) | | 🐚 **One-line installer** | `curl -fsSL …/install.sh \| sh` | [Installer](docs/INSTALLATION.md#linux--macos--one-line-installer) | | 🪟 **Windows** | zip from [Releases](https://github.com/bankyresearch/duxxdb/releases) | [Windows guide](docs/INSTALLATION.md#windows--zip-download-manual) | | ☸ **Kubernetes** | `kubectl apply -f packaging/k8s/duxxdb.yaml` | [k8s guide](docs/INSTALLATION.md#kubernetes) | | 🐍 **Python** | `pip install …` (locally built wheel) | [Python](docs/INSTALLATION.md#embed--python-wheel) | | 📦 **Node / TS** | `npm install …` (locally built `.node`) | [Node](docs/INSTALLATION.md#embed--node--typescript) | | 🦀 **Rust crate** | `cargo add duxx-memory --git …` | [Cargo](docs/INSTALLATION.md#embed--rust-crate) | | 📐 **From source** | `cargo build --release --workspace` | [Source](docs/INSTALLATION.md#from-source-any-os) | → Full per-platform install + uninstall + service-control cheatsheet: **[docs/INSTALLATION.md](docs/INSTALLATION.md)**. --- ## Quickstart #### From a redis client (any language) ```bash redis-cli -p 6379 127.0.0.1:6379> REMEMBER alice "I want a refund for order #9910" (integer) 1 127.0.0.1:6379> RECALL alice "refund" 5 1) 1) (integer) 1 2) "0.032787" 3) "I want a refund for order #9910" 127.0.0.1:6379> PSUBSCRIBE memory.* ``` #### From Python ```python import duxxdb store = duxxdb.MemoryStore(dim=32) store.remember(key="alice", text="I lost my wallet") hits = store.recall(key="alice", query="wallet", k=5) for hit in hits: print(hit.id, hit.score, hit.text) ``` #### From Node / TypeScript ```ts import { MemoryStore } from "duxxdb"; const store = new MemoryStore({ dim: 32 }); store.remember("alice", "I lost my wallet"); console.log(store.recall("alice", "wallet", 5)); ``` #### Embedded in a Rust app ```rust use duxx_memory::MemoryStore; use duxx_embed::HashEmbedder; use std::sync::Arc; let store = MemoryStore::new(Arc::new(HashEmbedder::new(32))); store.remember("alice", "I lost my wallet", None)?; let hits = store.recall("alice", "wallet", 5)?; ``` #### Over gRPC (any tonic / grpcio / grpc-go client) ```bash grpcurl -plaintext -d '{"key":"alice","text":"I lost my wallet"}' \ localhost:50051 duxx.v1.Duxx/Remember grpcurl -plaintext -d '{"key":"alice","query":"wallet","k":5}' \ localhost:50051 duxx.v1.Duxx/Recall ``` #### Over MCP (Claude Desktop / Cline / any MCP agent) ```jsonc // claude_desktop_config.json { "mcpServers": { "duxxdb": { "command": "/usr/local/bin/duxx-mcp" } } } ``` → Full per-surface walkthrough: **[docs/USER_GUIDE.md](docs/USER_GUIDE.md)**. → Wiring into a real agent (chatbot / voice bot / migration): **[docs/INTEGRATION_GUIDE.md](docs/INTEGRATION_GUIDE.md)**. --- ## Architecture ``` AGENT (chatbot / voice bot / autonomous) │ ┌─────────────────────┴─────────────────────┐ ▼ ▼ embedded ┌───────────────┐ duxx-memory (lib) │ duxx-server │ RESP TCP │ ├───────────────┤ ▼ │ duxx-grpc │ gRPC + streaming Python / Node bindings ├───────────────┤ │ duxx-mcp │ stdio JSON-RPC └───────┬───────┘ │ ┌───────────────────────┼──────────────────┐ │ │ │ ▼ ▼ ▼ duxx-memory duxx-reactive duxx-coldtier MEMORY / TOOL_CACHE (ChangeBus) (Parquet) / SESSION + decay │ ┌─────────────┼─────────────┐ ▼ ▼ ▼ duxx-query duxx-index duxx-storage duxx-embed (RRF + tantivy + redb / hash / OpenAI / hybrid) hnsw_rs memory Cohere ``` 21 Rust crates, all Apache 2.0. Full design: **[docs/ARCHITECTURE.md](docs/ARCHITECTURE.md)**. --- ## Benchmarks Hybrid recall, k = 10, single thread, debug release, toy 32-dim embedder. Reproduce: `cargo bench -p duxx-bench`. | Corpus | Median recall | Throughput | vs 10 ms target | |---:|---:|---:|---:| | 100 docs | **123 µs** | 8.1 k QPS | 81× headroom | | 1 000 docs | **166 µs** | 6.0 k QPS | 60× headroom | | 10 000 docs | **373 µs** | 2.7 k QPS | 27× headroom | For a cross-system comparison — retrieval quality (recall@k / nDCG@k vs an exact ground truth), latency, and throughput under concurrent load, with DuxxDB run **disk-backed** against disk-backed Redis / Qdrant / pgvector — run the harness (`bench/comparative/run.sh`) and read the methodology in [`docs/BENCHMARKS.md`](docs/BENCHMARKS.md). We don't publish an unqualified "Nx faster" headline: any speed claim carries its workload, dimensions, hardware, and the peer's configuration. --- ## How DuxxDB compares | | Vector | BM25 | Structured | Embedded | Server | Reactive | Agent prims | MCP | |---|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:| | **Redis** | ✗ | ✗ | ⚠ | ✗ | ✅ | ✅ | ✗ | ✗ | | **Valkey** | ⚠ (mod) | ⚠ (mod) | ⚠ | ✗ | ✅ | ✅ | ✗ | ✗ | | **Qdrant** | ✅ | ⚠ | ⚠ | ✗ | ✅ | ⚠ | ✗ | ✗ | | **Pinecone** | ✅ | ✗ | ⚠ | ✗ | ✅ | ✗ | ✗ | ✗ | | **Milvus** | ✅ | ⚠ | ✅ | ✗ | ✅ | ✗ | ✗ | ✗ | | **Weaviate** | ✅ | ✅ | ✅ | ✗ | ✅ | ✗ | ✗ | ✗ | | **pgvector** | ✅ | ⚠ | ✅ | ✗ | ✅ | ⚠ | ✗ | ✗ | | **LanceDB** | ✅ | ⚠ | ✅ | ✅ | ⚠ | ✗ | ✗ | ✗ | | **ChromaDB** | ✅ | ✗ | ⚠ | ✅ | ✅ | ✗ | ⚠ | ✗ | | **DuxxDB** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | Full take with "when each one is the right choice" + coexistence patterns: [docs/PROJECT_OVERVIEW.md § 8](docs/PROJECT_OVERVIEW.md#8-competitive-landscape). --- ## Documentation | Doc | Use it when | |---|---| | [INSTALLATION.md](docs/INSTALLATION.md) | Picking install method per OS | | [USER_GUIDE.md](docs/USER_GUIDE.md) | Writing client code in any of 6 languages | | [INTEGRATION_GUIDE.md](docs/INTEGRATION_GUIDE.md) | Wiring DuxxDB into a chatbot, voice bot, or autonomous agent (with diagrams) | | [DUXX_STACK_INTEGRATION.md](docs/DUXX_STACK_INTEGRATION.md) | How **duxx-ai** (Python framework) plugs into DuxxDB — backend protocol, API contracts, deployment shapes | | [DRAGONFLY_COMPARISON.md](docs/DRAGONFLY_COMPARISON.md) | Enterprise reliability roadmap inspired by DragonflyDB's operational model | | [FAQ.md](docs/FAQ.md) | "Is DuxxDB right for me?" — when to use, when not to, framework support, migration | | [PROJECT_OVERVIEW.md](docs/PROJECT_OVERVIEW.md) | Pitching DuxxDB to a teammate / making a build-vs-buy call | | [ARCHITECTURE.md](docs/ARCHITECTURE.md) | Understanding the internals before contributing | | [ROADMAP.md](docs/ROADMAP.md) | Knowing what's shipped vs. planned | | [CONTRIBUTING.md](CONTRIBUTING.md) | Opening your first PR | --- ## Repository layout ``` duxxdb/ ├── Cargo.toml workspace root (21 crates) ├── Dockerfile multi-stage, multi-arch image ├── packaging/ Postgres-style installers │ ├── docker/ docker-compose.yml + .env.example + prometheus.yml │ ├── systemd/ duxxdb.service + duxx.env │ ├── debian/ .deb postinst / prerm / postrm │ ├── homebrew/ duxxdb.rb formula │ ├── k8s/ StatefulSet + Service + ConfigMap │ └── scripts/install.sh curl-pipe installer ├── crates/ │ ├── duxx-core/ Schema · Value · Error │ ├── duxx-storage/ Storage trait + redb / memory backends │ ├── duxx-index/ tantivy BM25 + hnsw_rs HNSW (+ compaction) │ ├── duxx-query/ RRF, hybrid recall │ ├── duxx-memory/ MEMORY / TOOL_CACHE / SESSION + decay + compact │ ├── duxx-reactive/ ChangeBus pub/sub │ ├── duxx-embed/ hash + OpenAI + Cohere embedders │ ├── duxx-trace/ agent trace store (Phase 7.1) │ ├── duxx-prompts/ versioned prompt registry (Phase 7.2) │ ├── duxx-datasets/ versioned eval datasets (Phase 7.3) │ ├── duxx-eval/ eval runs + regressions (Phase 7.4) │ ├── duxx-replay/ deterministic agent replay (Phase 7.5) │ ├── duxx-cost/ token + cost ledger (Phase 7.6) │ ├── duxx-docs/ document-intelligence layer (ingest→cite) │ ├── duxx-token/ signed short-lived credentials (JWT) │ ├── duxx-server/ RESP2/3 daemon + auth + Prometheus │ ├── duxx-mcp/ MCP stdio JSON-RPC ser … ## Source & license This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [bankyresearch](https://github.com/bankyresearch) - **Source:** [bankyresearch/duxxdb](https://github.com/bankyresearch/duxxdb) - **License:** Apache-2.0 - **Homepage:** https://www.duxx.ai/duxxdb Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** yes - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** yes - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: flagged — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/mcp-bankyresearch-duxxdb - Seller: https://agentstack.voostack.com/s/bankyresearch - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.