Connected Apps for Agent Context

Aug 2026What is the difference between an MCP and a CLI?This post introduces the Model Context Protocol (MCP) as a transport for AI agents, contrasting it with the Command Line Interface (CLI). It details how MCP enables structured tool discovery and access with built-in access controls, making it suitable for business users and shared infrastructure. The post explains the token cost implications of MCP's pre-loading of tool definitions and outlines a decision framework for choosing between CLI and MCP based on user type and work environment. It also highlights how Hex integrates both CLI (Hex CLI) and MCP (Hex MCP server) to provide flexible agent connectivity for different use cases, emphasizing the importance of a shared context layer for consistent analysis.

Aug 2026Context in Analytics: Turning Raw Data Into Actionable InsightThis post details the evolution of Hex's approach to 'contextual analytics,' emphasizing the critical role of business meaning, lineage, quality signals, and policy in making data trustworthy for both human analysts and AI systems. It highlights how tool sprawl destroys context and how AI makes context the bottleneck due to its inability to inherently understand business-specific definitions. The post introduces the concept of a centralized context layer as a hub-and-spoke architecture, where semantic authoring and governed context directly feed into AI models. It explains how this context compounds over time, making future answers more accurate and grounded. The post also provides practical advice on getting started with building context, starting with endorsed tables and warehouse descriptions and layering in semantic models and governance practices as needed. The core technical contribution is the articulation of a system design that prioritizes business meaning as infrastructure to enable scalable and trustworthy AI analytics.

Jul 2026How the best data teams obsess over contextThis post details how Hex customers (Ramp, Underdog, Hover) are actively managing and evolving the context provided to their AI analytics agents. It highlights the use of dbt integrations for syncing model descriptions and domain docs, GitHub Actions for managing context guides, and API syncs for endorsing trusted models. The post also emphasizes the role of Context Studio in observing agent interactions, identifying context gaps through warnings and suggestions, and programmatically applying fixes back to dbt via the CLI and pull requests. This iterative process of leveraging user interactions to improve agent accuracy and trustworthiness is a key technical contribution.

Jul 2026How to Improve Agent Efficiency Through Context EngineeringThis post details context engineering for analytics agents, focusing on improving efficiency and accuracy. It introduces four techniques: retrieval (warehouse metadata, endorsed tables), memory (previous analyses, published notebooks), institutional knowledge (workspace guides), and codified metrics (semantic models). It also addresses context rot through ground truth testing, trajectory evaluation, conversation-length trends, and token efficiency tracking, highlighting Hex's Context Studio for surfacing user questions and recommending context improvements. The post emphasizes that governed context moves accuracy more than model choice.

Jun 2026Hex vs. Claude: Where Individual AI Analysis Meets Team AnalyticsThis post details how Hex integrates with AI assistants like Claude to provide governed answers anchored in shared definitions, contrasting this with the 'single-player' model of standalone AI assistants. It highlights the evolution of Hex's agent capabilities by emphasizing the creation of a shared, governed environment where individual AI analysis compounds into team value. The technical contribution lies in the architectural approach to bridging individual AI tools (like Claude Code) with a collaborative analytics platform (Hex). This involves using Hex's MCP server and CLI to publish findings directly into Hex as live data apps, and leveraging the Hex connector in Claude to access governed answers. The core technical concept is the use of 'trusted context' built from endorsed tables, workspace rules, and semantic models to ensure consistency and governance across different interaction points (Threads, notebooks, Slack, MCP). This contrasts with AI assistants that reconstruct business logic from raw tables on every request, leading to metric chaos. The post also touches on the validation bottleneck and how Hex's Context Studio provides visibility into agent usage and context gaps.

Jun 2026Introducing Agent TasksIntroduced Agent Tasks, a feature that allows scheduling of Hex Agent prompts for automated execution and delivery of results via Slack or email. This enables recurring question answering, proactive metric monitoring with root cause analysis, and consolidated account updates by integrating with various data sources and learning from previous runs. The feature aims to provide the consistency of dashboards with the reasoning capabilities of an agent.

Jun 2026Bring the web to your dataThis post introduces web search capabilities for the Hex Agent, integrated as a native tool. It enables two primary functions: open-ended web search to answer questions requiring external context (e.g., industry benchmarks) and URL extraction to read and bring back relevant information from web pages. This functionality is available via Slack or the MCP and was developed in collaboration with Parallel.

May 2026Connect Hex to the rest of your stackThis post details the integration of external tools (Notion, Linear, and any tool exposing an MCP server) with the Hex agent. It explains how this extends the agent's context beyond warehouse data to include meeting notes, tickets, and planning documents. The agent can now reason across this combined context to provide more informed answers and drive actions across the user's stack. It highlights the use of individual authentication for connected tools to respect existing permissions and emphasizes the benefit of cross-tool reasoning for richer insights.

May 2026Tactics for credit efficiency in HexThis post details tactics for optimizing AI credit efficiency within Hex. It explains the effort-based AI pricing model and provides three key tactics: 1. Building out context through table metadata, endorsements, and guides, including programmatic syncing from GitHub and leveraging existing Hex projects as context. It also emphasizes keeping guides concise and using Context Studio to monitor usage and identify areas for improvement. 2. Educating users on smarter prompting, including referencing existing assets like Collections and Data Apps, and reserving high-cost models for complex tasks. 3. Utilizing admin controls for managing and planning spend, such as setting thoughtful add-on caps and per-user credit allocations.

May 2026Your code knows things your warehouse doesn'tThis post introduces the capability for Hex Agents to connect to and analyze Git repositories (dbt, application code). This allows the agent to understand the logic behind data transformations and application features, providing context that was previously only accessible to data teams. This enhances the agent's ability to answer questions about data lineage, filtering, and feature implementation, thereby improving self-service analytics for a wider range of users.

May 2026AI Analytics Adoption Risks: 7 Pitfalls to AvoidThis post details seven pitfalls to avoid in AI analytics adoption, focusing on governance, context, and fragmentation rather than model quality. It highlights risks such as ungoverned AI tool spread leading to metric drift and security incidents, the danger of trusting AI answers without verifiable context (citing the Spider 2.0 benchmark's findings on text-to-SQL accuracy collapse), the fragmentation problem caused by multiple AI point solutions, and the unrealistic expectation of 'plug in and go' self-serve AI without proper context and governance. It also discusses treating semantic layers as prerequisites instead of maturity paths, skipping feedback loops between AI output quality and context, and prioritizing model chasing over data quality. The post advocates for unified workspaces and curated context, such as endorsed tables and semantic models, to address these issues.

May 2026Building Topic Discovery into the Hex Context AgentThis post details the development and evolution of topic discovery for the Hex Context Agent. Initially, a complex ML pipeline involving embeddings and HDBSCAN was considered but deemed too infrastructure-heavy and not suitable for recurring product use. The team then pivoted to a simpler, LLM-centric approach. The key innovation was splitting the process into two stages: topic discovery (LLM proposes topics and examples) and batch classification (LLM assigns threads to discovered topics). This approach, leveraging LLMs for both clustering and naming, significantly reduced infrastructure requirements and enabled a more scalable and maintainable solution. The system now runs incrementally, classifying new threads and periodically re-discovering topics to capture emerging trends. Limitations include a flat topic structure, with future work planned for hierarchical clustering and proactive context injection based on identified topics.

Apr 2026Why is no one talking about context exhaust?This post introduces the concept of 'Context Exhaust,' which is the implicit context generated by user interactions with AI agents and software. It describes how this exhaust, when captured and refined into semantic models, guides, and memories, bootstraps knowledge for accurate and trusted answers, creating a compounding feedback loop. Hex enables this by allowing data teams to sync external context sources (like dbt repos), commit agent interactions to a context repo, observe agent responses for potential inaccuracies, and receive suggestions for improvement. The net effect is a system that gets smarter with use, fostering collective learning and iterative evolution of organizational knowledge.

Apr 2026Introducing Context SuggestionsThis post introduces Context Suggestions, an automated system for improving agent context. It details a 'self-learning flywheel' where a Review Agent analyzes user Threads to extract summaries, topic tags, and warnings (context gaps, data limitations, user doubt). These insights are synthesized into 'Suggestions' for data teams to update Guides, semantic models, or metadata. The system aims to make Hex smarter over time by leveraging user interactions to continuously enhance agent knowledge.

Apr 2026AI Sprawl: Causes, Risks & 7-Step Control PlanThis post details the causes and risks of AI sprawl, including metric trust collapse and the hidden tax on data teams, stemming from disconnected AI tools and a lack of centralized, trusted context. It proposes a 7-step plan to control AI sprawl, emphasizing the importance of inventorying existing tools, centralizing metadata over compute, and defining high-value metrics in a semantic layer. The plan advocates for incremental adoption of governance, starting with the most critical definitions, to ensure AI-generated answers are consistent and trustworthy.

Apr 2026Context Engineering Best Practices: A Hands-On GuideThis post introduces the concept of 'context engineering' for AI analytics, defining it as the practice of building structured environments to make AI answers accurate and reproducible. It outlines four layers of context: endorsed tables and warehouse descriptions, workspace rules, semantic models, and observability/feedback loops. The post details how each layer addresses specific failure modes of AI agents, such as metric inconsistency, join proliferation, and confidently wrong answers. It emphasizes starting with endorsed tables and descriptions as the most immediate impactful step, and then progressively investing in workspace rules and semantic models based on observability data from Hex's Context Studio.

Apr 2026Context Engineering vs Prompt Engineering: What Data Teams NeedThis post introduces the concept of 'context engineering' as a necessary evolution beyond 'prompt engineering' for data teams using AI. It details the components of context engineering: endorsed tables, warehouse schema and table descriptions, workspace rules, semantic models, conversation history, and observability signals. It also outlines a practical, phased approach for data teams to implement these context layers, starting with endorsing tables and adding descriptions, then moving to workspace rules and semantic models. The post argues that context engineering is essential for ensuring trusted, consistent, and auditable analytical answers from AI, especially at enterprise scale, and contrasts this with the limitations of prompt engineering for complex data tasks. It also discusses how this approach addresses the problem of 'shadow AI' by making the governed path faster and more reliable.

Mar 2026Projects as Context: Hex Now Learns From the Work You Already BuiltThis post introduces the capability for Hex agents to use existing Hex projects as context. Previously, agents could analyze code repositories. Now, agents can search and reference completed Hex projects (analyses, dashboards, logic) to answer user questions. This includes surfacing existing trusted work, using it as a reference for new analyses, and improving documentation by allowing agents to read project context directly. The system respects 'Endorsed Mode' for trusted assets and can be explicitly directed via @-mentions or URLs.

Feb 2026Context-Aware AI in Analytics: Useful Answers vs Confident GuessesThis post details the evolution of Hex's AI capabilities towards context-aware AI in analytics. It explains how combining LLMs with specific data, definitions, and business logic (schema, metric definitions, permissions, workspace history) leads to more accurate and trustworthy answers compared to generic chatbots. It highlights the application of RAG principles with richer, structured context (live schema metadata, governed metric definitions, query history, team-specific rules) and contrasts 'bolt-on' AI with AI native to the analytical workflow. The post emphasizes the importance of governed semantic layers and introduces Context Studio for observability into context breakdown and quality issues. It also discusses trade-offs like context accuracy, freshness, predictability, and user control, advocating for visible AI reasoning and the ability to inspect/override decisions.

Feb 2026Conversational Analytics: How to Get Data Answers in Plain EnglishThis post details the technical architecture and trustworthiness mechanisms of conversational analytics, focusing on how natural language questions are translated into database queries. It highlights the importance of schema linking, business logic integration, and semantic models for generating accurate and actionable SQL. The post contrasts this with generic chatbots, emphasizing the need for direct database connection and context awareness. It also discusses the evolution from simple chatbots to integrated AI analytics platforms where conversational queries can lead to inspectable notebooks and dashboards.

Jan 2026The data team's guide to generative AI analyticsThis post details how generative AI analytics works by translating natural language to code, speeding up the analysis loop, opening access to non-technical users, maintaining governance through semantic models and editable outputs, and facilitating collaboration. It also outlines how analytics engineers, data scientists, and business users can leverage these capabilities for pipeline development, exploratory analysis, and self-serve data exploration.

Jan 2026Introducing Context Studio — Curate trusted answers for your whole team, wherever they’re working.This post introduces Context Studio, a new system for observing, diagnosing, curating, and testing AI agent context. Key technical contributions include: an observability dashboard for tracking agent adoption and user questions; a Thread inspector for diagnosing agent confusion by analyzing requests, responses, topics, and data sources; a multi-layered context management system with endorsements, semantic models, and workspace guides; and a context workbench for staging, testing, and deploying changes to workspace guides and semantic models with version control and attribution logs.

Dec 2025How the best data teams use the Notebook AgentThis post details how companies like Ramp, Figma, and Notion are using the Notebook Agent for analytics. It highlights how the agent assists in planning and building analyses, visualizing metrics, and summarizing results, reducing the need to switch between tools. Specific use cases include tracking feature performance, transforming survey responses into insights, building reusable components, and accelerating report development. The agent is described as a collaborator for tasks like predictive modeling and exploring new domains. The post emphasizes the agent's ability to explain its thinking and show its work, building trust in its outputs.

Dec 2025What is natural language BI?This post details the mechanics of natural language BI, explaining how it uses NLP and a text-to-SQL pipeline to translate natural language questions into executable database queries. It highlights the role of the semantic layer in mapping business terms to database schema and ensuring consistency. The post also describes conversational systems that maintain context across sessions, enabling follow-up questions and iterative exploration. Use cases include ad hoc investigations, self-service reporting, meeting preparation, real-time decision-making, and collaborative team exploration. It emphasizes the importance of preparing data, focusing on trusted tables, and implementing data governance for successful adoption.

Dec 2025What is AI for data analysis?This post explains the concept of AI data analysis, detailing how natural language processing, ensemble machine learning, governance frameworks, and semantic layers work together to automate data analysis workflows. It contrasts AI-assisted analysis with traditional approaches, highlighting Hex's integrated platform that supports AI for data scientists (notebook agents), business users (natural language interfaces), and exploration of existing artifacts. Key components discussed include transformer-based language models for understanding queries, ensemble methods like Random Forests for pattern detection, governance frameworks for access control and business logic, semantic layers for consistent metric definitions, and model explainability techniques. The post also evaluates current AI analytics platforms, positioning Hex as a solution that offers AI at every level with connected experiences, shared context, governance, and observability, extending AI through integrations like Slack and MCP.

Nov 2025How context engineering makes or breaks AI-driven self-serviceThis post details how context engineering, encompassing warehouse metadata, endorsements, rules files, and semantic models, is crucial for grounding AI agents in business knowledge, enabling trustworthy self-service analytics and transforming the data team's role from query mechanics to context engineers. It emphasizes a progressive rollout of context curation and a hybrid architecture combining deterministic semantic models with flexible rules and metadata.

Oct 2025Hex Agent, everywhere: Introducing Slack integration and MCP supportIntroduced a native Slack integration for the Hex Agent, allowing users to ask data questions directly in Slack channels. The agent responds with answers, explanations, charts, and links to Hex projects. Also introduced an MCP server to enable AI assistants and custom tools to access Hex context and initiate analyses, facilitating integration with code editors and other AI tools. This expands the accessibility of Hex's agentic capabilities beyond the Hex workspace.

Oct 2025Improving the Notebook Agent: Error ClusteringThis post details the application of cluster analysis (specifically K-Means on OpenAI embeddings) to analyze errored tool calls from Hex's Notebook Agent. It describes the process of embedding agent thinking messages, subdividing by tool, clustering, and using an LLM to label clusters. Key findings include identifying common error patterns for specific tools (e.g., RunSQL, CreateChartCell), the impact of timeouts leading to proliferation of simpler cells, and the critical importance of clear error messages for graceful agent recovery. The post also outlines specific improvements made based on these findings, such as prompt enhancements for Dataframe SQL, introducing a 'notebook DAG' tool to improve cell editing over creation, and modifying timeout handling to involve user consultation.

Oct 2025Why Analytics Agents Break DifferentlyThis post details the challenges and solutions for context management in analytics agents, specifically highlighting how data objects break differently than code objects due to their unpredictable size and lack of inherent abstraction. Key contributions include the concept of structuring context as a computational graph (notebook map) rather than a file tree, implementing fixed token boundaries for predictable agent behavior and cost, and the importance of explicit truncation messaging to the agent. It emphasizes that analytics agents require discovered structure, unlike coding agents which rely on predetermined abstractions.

Oct 2025Bitter lessons building AI productsThis post details lessons learned in building AI products, emphasizing the importance of adapting roadmaps to evolving model capabilities rather than relying on clever engineering to overcome model deficiencies. It highlights the failure of an early Notebook Agent due to model limitations and the "pretty good" trap with visualization generation, where clever optimizations were outpaced by new model capabilities like tool calling. The team now prioritizes early customer validation, spotting capability shifts, and faster project killing. The post advocates for building AI features that leverage future model improvements rather than working around current limitations.

Oct 2025What is agentic analytics?This post defines agentic analytics as a paradigm shift driven by AI agents capable of step-by-step reasoning, tool utilization, and planning. It outlines three key dimensions of change: depth of insight (moving beyond single-layer dashboards to complex, multi-step analysis), breadth of reach (democratizing data access for non-data professionals and enabling agents to interact via various channels like Slack), and speed of insight (accelerating the full path to insight through asynchronous patterns and proactive surfacing). The post emphasizes that this shift is not product-specific but a fundamental change in how AI agents can accomplish analytics tasks.

Oct 2025Lessons from Ramp: Running AI-powered data teamsThis post details how Ramp's data team uses Hex's Notebook Agent to accelerate analysis and share insights. It highlights the use of the agent for building analysis frameworks and visualizations rapidly, accelerating exploratory analysis by testing hypotheses directly within the notebook, and synthesizing findings for quick distribution to non-technical audiences. A key aspect is the emphasis on providing context to the AI agent through a well-documented semantic layer, including company background, product details, data infrastructure, and endorsed tables. The post also touches on the use of the agent by quasi-technical users and the necessity of evals for ensuring reliability and accuracy of LLM outputs.

Sep 2025AI: because your team has better things to do than writing boilerplate SQLThis post details how the Hex data team uses AI, specifically the Notebook Agent, to automate repetitive tasks like drafting dbt documentation and bootstrapping analyses, and to unlock new insights by analyzing unstructured data. It emphasizes a pragmatic approach of testing AI tools liberally, time-boxing experiments, keeping humans in the loop, and ensuring AI-driven work influences business outcomes. The post also highlights current limitations of AI in complex data modeling and bug identification, and the ongoing evolution of AI capabilities in areas like visualization and reporting. The core contribution is the team's methodology for applying AI to data work to increase speed and impact, rather than a specific new technical feature.

Sep 2025How to Work with Hex's Notebook Agent: A Prompting GuideThis post details how to effectively prompt Hex's Notebook Agent, providing mental models and copy-paste templates for structured and conversational prompting. It emphasizes scoping context by tagging tables and cells, specifying analysis methods, and providing business context for impact. The post also introduces the concept of "meta prompting" where an agent's plan is fed back for execution and highlights the agent's key capabilities: agentic search, building a plan, executing analysis, and summarizing results.

Sep 2025How Hex's Notebook Agent changed pricing analysis with AI-powered data analysisThis post details the introduction and application of the Notebook Agent within Hex. The agent is an AI assistant embedded in the Notebook experience that can write SQL/Python, pull data, and build charts from natural language prompts. The author used it for pricing analysis, demonstrating its ability to handle complex joins and window functions. The agent's output was validated by the data team, highlighting its accuracy and speed. The post emphasizes the agent's role in accelerating analysis and making complex tasks more approachable for users without deep SQL expertise.

Aug 2025Introducing the Notebook AgentIntroduces the Notebook Agent, an AI assistant integrated into the Hex Notebook environment. The agent can plan analyses, search data, create and chain SQL/Python cells, visualize data with pivot and chart cells, and summarize findings. It supports mentioning cells/tables for context, autosaves agent-initiated work for version restoration, and provides cell-by-cell diff views for reviewing and editing generated code. The post also mentions upcoming integration with semantic models for governed context.

Aug 2024The future of analytics is multi-modalThis post articulates the vision for a multi-modal analytics future, where different languages (SQL, Python, R) and interfaces (spreadsheets, AI, chart builders) are treated as modalities of working with data. It highlights Hex's 'cells' system as the foundation for this multi-modality, allowing users to combine code, no-code UI elements, and AI agents within a single integrated workflow. The post emphasizes that this approach aims to solve the fragmentation of data workflows by providing a unified environment for diverse data tasks, contrasting it with current mono-modal tools. It references existing Hex features like the Notebook Agent and Threads as examples of this multi-modal approach.

What this trail covers