DuckDB’s Quack Protocol Enables Client-Server with Multiple Concurrent Writers
DuckDB has released the Quack remote protocol, enabling DuckDB instances to communicate in a client-server setup with support for multiple concurrent writers. The protocol is designed to be simple, fast, and builds on proven technologies like HTTP.
While DuckDB’s in-process architecture excels for single-process data science workloads, concurrent writes to the same database file from multiple processes were a pain point. Common workarounds included custom RPC solutions, Arrow Flight SQL, MotherDuck’s proprietary protocol, or even switching to PostgreSQL (optionally running DuckDB via pg_duckdb). Quack is DuckDB’s official first-party solution.
How It Works
Both DuckDB instances need the Quack extension, currently available in the core_nightl (nightly) build. The protocol is symmetric: any DuckDB instance can act as client or server. The server exposes an HTTP endpoint; the client connects to it. No special server binary is required—just a running DuckDB process with the extension loaded.
Key design choices:
- Protocol is built on HTTP, making it firewall-friendly and debuggable with standard tools.
- Supports bulk operations and small transactions with low latency.
- No legacy constraints—designed from scratch in 2026, learning from Arrow Flight SQL and others.
Use Cases
- Multiple telemetry-collecting processes inserting into the same DuckDB database.
- A dashboard querying the same tables concurrently without file-locking issues.
- Enabling DuckDB in traditional client-server deployments without external middleware.
This release marks a shift for DuckDB, which previously emphasized its in-process nature. The team acknowledges user demand as the driver: “We see DuckDB as a universal data wrangling tool. If this means having a client-server protocol in addition to the in-process capabilities – fine.”
Current status: Extension is available in the nightly build. Expect a stable release in the coming months.
For a deep dive into the protocol internals, including the full research paper reference and a step-by-step setup guide, read the original announcement.
📖 Read the full source: HN AI Agents
👀 See Also

The Bottleneck in Parallel AI Agents: Human Approval Queue Bottleself
A developer running parallel Claude Code agents describes the 'bottleself' — the point where parallelism stops adding output and starts creating a backlog of human approvals. Their solution: a planner that decomposes goals into subtasks, spawns agents, and only pings on unresolved decisions.

pop-pay MCP server adds payment guardrails for Claude Code agents
pop-pay is an MCP server that lets Claude Code agents handle purchases without exposing credit card numbers. It uses CDP injection to place virtual card credentials directly into payment iframes, with Claude only receiving masked confirmation numbers.

VoidLLM: Zero-Knowledge Proxy for Ollama and vLLM with Team Access Control
VoidLLM is a proxy that sits between applications and local LLM servers like Ollama and vLLM, adding organization/team access control, API key management, usage tracking, and rate limiting without viewing prompts. It has <2ms proxy overhead and works with OpenAI-compatible SDKs.

mnemos: A Persistent Memory Layer for AI Coding Agents (Go, MCP-Native, No Python)
mnemos is a Go-based MCP-native memory layer for AI coding agents. The author built a verifier to measure lift: +40% aggregate on read-side scenarios, but only 53% write-side capture rate after iterative fixes.