Spice AI

Using Spice as an MCP gateway and model router for a personal agent platform 🤖 This new demo covers two integrations patterns for deploying Spice as a data platform behind OpenClaw. Either use Spice as an MCP gateway with your own LLM, or point directly at Spice's model routing and let it run the tool loop internally. Both patterns offer: 1. Federated and accelerated data: One SQL surface across any data source. 2. MCP gateway: SQL tools, a local stdio MCP, and a remote HTTP MCP unified behind one endpoint. 3. Model routing: Three models behind the same point, and the Tool Registry keeps per-turn token cost flat as the catalog grows, regardless of how many tools are registered. Check it out! Code and walkthrough in the repo: https://hubs.ly/Q04jpfDP0

Analysts querying a data warehouse 9-5 is a very different problem from hundreds or thousands of agents running against every feature or product 24/7. This is why we believe every agent needs its own data stack.

Meet with the Spice team at the Databricks Data + AI Summit 🤝 You can find us at booth 107 throughout the week, but we'd also love to get together for a coffee or meal. Book a time here: https://hubs.ly/Q04j2Xhg0 #DatabricksPartner

OpenClaw is powerful because it removes all guardrails and runs anywhere. At DeepLearning.AI's AI Dev 2026, we demoed what an OpenClaw implementation might look like in an enterprise environment with additional security and human oversight in place. The demo builds out an SRE agent that detects and diagnoses production incidents across Postgres, GitHub runbooks, and HTTP APIs, all through a single Spice SQL endpoint. The agent never holds direct credentials and every query is traced. The agent handles the investigation, Spice manages data access, and a human makes the final call over Slack. Check out the walk-through & demo recording: https://hubs.ly/Q04hWZjn0

OpenClaw is powerful because it removes all guardrails and runs anywhere. At DeepLearning.AI's AI Dev 2026, we demoed what an OpenClaw implementation might look like in an enterprise environment with additional security and human oversight in place. The demo builds out an SRE agent that detects and diagnoses production incidents across Postgres, GitHub runbooks, and HTTP APIs, all through a single Spice SQL endpoint. The agent never holds direct credentials and every query is traced. The agent handles the investigation, Spice manages data access, and a human makes the final call over Slack. Check out the walk-through & demo recording: https://hubs.ly/Q04hWZjn0

RRF (Reciprocal Rank Fusion) merges full-text and semantic results by rank rather than raw score. It makes combining search methods that score on different scales much more reliable. RRF in Spice enables you to layer on per-query rank weights, recency decay (exponential or linear), filter pushdown into subqueries before ranking, and an optional rerank step using a dedicated reranker model or LLM. A few examples below! The Spice RRF cookbook will help get you started: https://hubs.ly/Q04hqR960

New to Spice? Check out this hands-on guide that introduces Spice from the ground up. The guide covers: • What Spice is • When to use it • How to use it Progressively develop your mental model of Spice by walking through the core concepts. ➡️: https://hubs.ly/Q04hbzDH0

Apache DataFusion Ballista takes DataFusion from single-node to multi-node. A scheduler coordinates across multiple executors, breaks queries into stages and partitions, and shuffles intermediate results between them. The architecture will feel familiar if you've used Spark, but Ballista is built in Rust, Arrow-native, and brings some significant performance improvements over Spark. Ballista docs: https://hubs.ly/Q04h3dBZ0

Three format-level problems Vortex solves for agents 🤖 Every engineering team is reasoning through what infrastructure is needed for agentic workloads that introduce far different requirements than traditional access patterns. There's a lot of surface area to this challenge, and one of the most fundamental is the underlying data format. Here are some of the patterns we've seen across teams building enterprise-scale agents, and how Vortex helps solve them. 🔷 Problem 1: Traditional formats weren't designed for random access Most columnar formats are optimized for full scans, but agents pull targeted subsets from large datasets. Formats like Parquet weren't built for that pattern, and memory pressure and latency compound with every agent query as a result. Vortex is designed for selective reads, so agents fetch exactly the rows and columns they need without scanning unnecessary data. Random access runs 100x faster than Parquet. 🔷 Problem 2: Single file architectures create contention under concurrency Production deployments involve hundreds or thousands of concurrent queries hitting the same datasets simultaneously. Single file formats serialize this access, and each query waits for others to complete. Vortex's multi-file architecture means agents can fetch context across terabytes in parallel without contention. Response times stay consistent. 🔷 Problem 3: Format conversion adds latency at every step Data pipelines generally need to move data through multiple formats before it reaches an agent. Vortex stores content, vectors, metadata, and structured data together. It is also natively compatible with Apache Arrow through zero-copy access, so data moves from storage to agent memory without serialization or conversion overhead. Vortex isn't a panacea for every use case, but for agent workloads that demand fast, concurrent, low-overhead data access, it's a powerful foundation. #vortex #data #arrow #parquet #spiceai

2 minute speed run of the historical arc of Spice, and what the Cluster-Sidecar architecture enables for data-intensive agents. Built on Apache DataFusion, Ballista, and Vortex.