Read time: ~8 minutes · Role: Analyst · Next: Work through the reading path below article by article, or jump straight to USER_GUIDE §2 and walk through a funnel analysis.
You may have seen this scenario
On Friday you spent two hours in AI asking about funnel and GMV metrics, and the numbers in the conversation looked reasonable. On Monday finance follows up: "Does the denominator include cancelled orders? How are refunds deducted?" You scroll back through the chat—only summaries, a few screenshots, and the JOIN and WHERE logic in the middle is impossible to explain.
The problem is usually not that the model isn't smart enough—it's that the analysis was left in chat bubbles: no Plan spelling out the metrics, no SQL tabs as evidence, no Report as a deliverable. Six months later, even you would struggle to reproduce that number.
What Lantide Data is (analyst view)
In one sentence: a local SQL analysis IDE with a built-in Agent copilot that can also connect controlled external Agents—no matter which conversation interface you work from, the metrics and data-pull logic should stay in SQL tabs and Markdown files you can review, not hidden in one-off Python cells or chat summaries.
| What you know | What Lantide does |
|---|---|
| Upload CSV → chat produces charts | Local DuckDB queries the files; the Agent shares the same table names as you |
| Metrics live in your head or in chat | Plan spells out the denominator, grain, and checkpoints; SQL tabs are the execution evidence |
| AI runs the analysis end to end | You press Execute for a formal run; the Planning stage only explores and revises the Plan |
| Conclusions pasted into Slack | Report lives in project files—annotatable and HTML-exportable |
It is not "ChatGPT that writes SQL," and not a "coding agent that edits repos"—it is an Agent execution environment built for analysis work: explore, draft, authorize execution, and deliver, all in one IDE.
Why it is worth using (instead of stitching tools together)
1. Metrics are debatable and rerunnable
How active users are defined, how two tables align on user_id, whether JOINs fan out—all written with FROM / JOIN / WHERE in SQL tabs and the Plan. Colleagues edit the same logic, not replay a conversation.
2. Analysis has stages, not a one-click black box
The Agent explores the schema and drafts a Plan first, and only runs Execute once you're satisfied with your annotations. For analysts this means: exploration can be fast, but a formal data pull needs a contract—the denominator, stage definitions, and limitations all have to be spelled out in the Report.
3. Quick and Project have clear roles
Column exploration and one-off validation → Quick Analysis (conversation + SQL tabs). Plan annotations, a formal Report, sign-off for managers → a focused project running Plan → Execute → Report. No more complaining "why is there no Approve & Execute button" in the wrong mode.
4. Advanced capabilities on demand
Cache, Source Run, statistical tools, Agent Memory—all available, but analysis correctness comes before pipeline form: there's no need to split steps just so the pipeline "looks complete"; a clear one-shot JOIN can run directly.
When it may not fit
- You only want one-off charts and never need to keep SQL or metric evidence
- Heavy custom Python / Notebook pipelines, and the team does not intend to treat SQL as the metric contract
- You expect the Agent to run reports end to end, fully automatically, with no human review
If you only need "quick Q&A," a general chat tool may be lighter; if you need to explain six months later how a number was produced, Lantide's design is a better fit.
Suggested reading path (getting started → practice → advanced)
Finish this article first, then read in order or jump around as needed.
Getting started: build a mental model (~15 minutes)
| # | Article | Overview | Status |
|---|---|---|---|
| 1 | Why analysis conclusions can't stay in chat | Chat lacks a contract, annotation binding, and locked delivery; when to upgrade to a project | Available |
| 2 | Quick Analysis vs Project Analysis | Which mode fits which tasks; how the Agent behaves when a project is focused | Available |
Practice: align with your daily workflow (~30 minutes + hands-on)
| # | Article | Overview | Status |
|---|---|---|---|
| 3 | Plan → annotate → Execute → Report | What to do at each stage; use Compare view for side-by-side checks during Execute / Report acceptance | Available |
| 4 | SQL-first: metrics in tabs | Data pulls in SQL; statistical tools on clean tables; vs chart-from-chat tools | Available |
| 5 | Cache and Source Run: how much to know | persist vs agent cache; when it's worth splitting out a persist tab | Available |
| — | Hands-on | USER_GUIDE §2 Quick Start (orders funnel) | Available |
Advanced: long-term collaboration and quality (~25 minutes)
| # | Article | Overview | Status |
|---|---|---|---|
| 6 | Statistical analysis tools in conversation | What activate_analysis can do; why the Agent asks for confirmation before running |
Available |
| 7 | Reference docs: where to put large references | Mapping tables / dictionaries vs Plan and Memory; Update Intro | Available |
| 8 | Skills: reusable analysis methods | Load with /; share a definition with external Agents; not a tool permission |
Available |
| 9 | Plan vs Agent Memory division of labor | One-off assumptions go in the Plan; cross-project metrics via Queued Knowledge approval | Available |
Design depth (optional)
| Resource | When to read |
|---|---|
| 1. Data analysis workflow in the Agent era | To understand the full Execute boundary and SQL-first methodology |
| 5. Unified query layer §8 | To understand the engineering detail behind materialization and Source Run |
| USER_GUIDE §5–15 | Operating queries, cache, Agent, Skills, and Memory |
| External Agent Integration | When the team wants to operate the same Lantide artifacts from Cursor, Codex, or Claude |
Trial checklist
- Read this article + the 2 getting-started articles
- Walk through USER_GUIDE §2 once: Plan → Execute → Report
- Try opening Compare view: cross-check the Plan during Execute, or cross-check the Plan when accepting the Report (see plan-execute-report §4–5)
- Leave one annotation on the Plan and confirm the Agent only changes the section you pointed at; after Resolve, try View changes and Archive
- (Optional) Build a small Reference mapping + Intro (see reference-docs-for-analysts)
- (Optional) Enable one personal Skill and load it with
/(see skills-for-analysts) - In Data → Cached, right-click View SQL and cross-check the metrics
- (If using an external Agent) Verify the Plan, SQL evidence, Report, and External MCP Activity inside Lantide
Other roles
- Product Manager introduction — adoption boundaries, team collaboration, the analysis quality bar
- Business & Operations introduction — review Plan, accept Report, sign off in the weekly meeting
- Platform Admin (Data Engineering) introduction — connections, workspace,
.lantidemigration - Learn Hub