Terminal-native Databricks agent

Databricks runbooks
in your terminal.

Databricks work too often lives in the UI: chasing failed runs, checking logs, tweaking compute, hunting permissions, verifying outputs, and explaining spend. Lakecode turns those UI workflows into repeatable terminal workflows with plan → execute → evidence.

terminal
$ lakecode /prove sales.orders
 
Collecting evidence for sales.orders...
Schema snapshot — 14 columns
Row count — 2,847,312
Null analysis — 0 critical nulls
Freshness — last update 23 min ago
Duplicate check — 0 duplicates on PK
Numeric stats — revenue within range
Sample rows — 5 rows attached
 
Health: PASS — Evidence pack saved to ./evidence/

What it does

Lakecode is a terminal-first Databricks engineering agent that helps you:

Debug & monitor jobs

  • Check job config + recent runs
  • Pull run output and logs
  • Analyze failures and suggest root causes with context (run metadata + query history)

Prove data outputs

  • Generate an evidence pack: schema, row counts, nulls, freshness, duplicates, numeric stats, sampling
  • Produce a plain-English health assessment you can paste into a PR or incident ticket

Understand spend & risk

  • Top spend breakdown from system billing data
  • Detect spikes and explain what changed vs baseline
  • Audit jobs for missing timeouts, SLA risk, failure patterns, and runtime variability

Deploy & execute

  • Deploy a local file into the workspace
  • Trigger a job run and poll until completion

How it works

Lakecode isn't "just a model + tools." It's a workflow engine that turns Databricks work into repeatable runbooks.

1

Builds workspace context automatically

On startup, lakecode generates a cached workspace profile (catalogs, schemas, table counts, warehouses, recent failures) so it can reason with the same context you'd normally hunt down in the UI.

2

Executes deterministic workflows first

Most steps are deterministic: SQL queries, API calls, log pulls, metadata inspection. These run through a step-based workflow engine with retries, abort handling, and progress updates.

3

Uses the LLM only where it adds value

Lakecode uses an LLM for interpreting error logs, ranking likely root causes, writing clean summaries, and turning raw outputs into actionable reports.

4

Stays safe by default

Plan/approval logic for complex or risky work, plus guardrails that prevent common mistakes (like treating function schemas as tables).

5

Produces evidence

For workflows like /prove, lakecode produces an evidence pack (schema, metrics, samples, assessments) so the outcome is shareable and reproducible.

Commands

Opinionated runbooks for real engineer workflows.

Debugging & Monitoring

/status job <id> Job config + last 5 runs
/logs <run-id> Run output + error analysis
/debug job <id> End-to-end diagnosis: config → failed run → errors → query history → root-cause summary
/queries job <id> SQL query history for the latest run
/queries run <run-id> SQL query history for a specific run

Data Quality

/prove <table> Evidence pack: schema, row counts, nulls, freshness, numeric stats, duplicates, samples + health assessment

Platform Ops

/audit jobs Scans jobs for SLA gaps, missing timeouts, runtime variability, failures; outputs risk-scored report
/cost top Top spend drivers from billing system tables
/cost spike Compare day vs baseline; identify likely drivers

Deployment

/deploy <local> <ws-path> Import file into workspace
/run job <id> Trigger job run + poll completion

Why lakecode

Workflows, not just tools

Lakecode ships opinionated runbooks for real engineer workflows: diagnose a failure, prove correctness, explain a cost spike, audit operational risk, deploy safely.

Proof beats "it ran"

Lakecode generates evidence you can attach to a PR, incident, or audit trail — not just a success/fail status.

Less context switching

Instead of bouncing between Jobs UI, cluster logs, UC permissions, SQL editor, and dashboards, lakecode collects the context and turns it into a single actionable report.

Roadmap

Available now

  • Workflow engine (deterministic + LLM steps)
  • Workspace profiling + conventions + skill injection
  • All core commands: /status, /logs, /debug, /queries, /prove, /audit, /cost, /deploy, /run
  • Plan mode + guardrails + tool truncation + conversation compaction

Coming next

  • Sessions + resume
  • Non-interactive mode for CI/scripts
  • Evidence pack scaffolding for every workflow
  • Watch mode (/watch job, /watch pipeline)
  • Drift detection + bundle adoption
  • Unity Catalog governance workflows
  • Optional model routing via Databricks AI Gateway
  • MLflow tracing for workflow/tool/model calls
  • MCP ecosystem management commands

Security & trust

Runs locally

Lakecode runs on your machine and connects to Databricks using your existing authentication (for example via the Databricks CLI profile).

Respects Unity Catalog permissions

Lakecode only sees what your Databricks identity can see. If you don't have access to a catalog, schema, or table, lakecode won't either.

Safe by default

Plan/approval logic for complex or risky operations, plus guardrails that prevent common high-cost or high-risk mistakes.

Evidence & auditability

Workflows produce reproducible artifacts so teams can share results instead of relying on screenshots and Slack messages.

Get started

Install lakecode and start running workflows in minutes.

terminal
$ pip install lakecode
$ lakecode
 
Welcome to lakecode. Profiling workspace...
Connected via profile: DEFAULT
Workspace profile cached (3 catalogs, 12 schemas, 847 tables)
 
Ready. Type a command or describe what you need.

FAQ

Is lakecode just a Databricks CLI wrapper?

No. Lakecode uses Databricks APIs and SQL to collect facts, but the value is the workflow engine: it chains steps, manages context, and produces a consistent report/evidence outcome.

Do I need to know Databricks APIs?

No. You describe the goal ("debug this job", "prove this table"), and lakecode executes the steps and explains results.

Does lakecode modify my workspace?

Some commands are read-only; others (like /deploy or /run job) perform actions. Lakecode is designed to be safe and explicit about what it's doing.

What models does it support?

Lakecode supports multiple model adapters (Anthropic + OpenAI today), and you can choose what to use based on your preferences and environment.

Is this a replacement for Databricks Assistant / AI Dev Kit?

Think of lakecode as terminal-native Databricks runbooks + mission control, optimized for reproducibility and evidence — whereas Assistant/AI Dev Kit are primarily tool/assistant building blocks.

Documentation

Overview

What lakecode is and what problem it solves — turning UI-driven Databricks work into repeatable, auditable workflows.

Read →

How it works

The step-based workflow engine: deterministic steps, LLM steps, data passing, retries, and abort handling.

Read →

Workflows & commands

Complete reference for all available commands across debugging, data quality, platform ops, and deployment.

Read →

Evidence packs

What evidence packs are, how /prove uses them today, and the roadmap for structured JSON outputs and replay plans.

Read →

Configuration

Conventions files, skills injection, default catalogs/schemas, naming patterns, and team standards.

Read →

Safety & plan mode

Plan/approval flow, guardrails, output truncation, and the roadmap for policy packs and org settings.

Read →

Docs: Overview

What is lakecode?

Lakecode is a terminal-native Databricks engineering agent. It turns UI-driven Databricks work — debugging runs, validating data, analyzing spend, auditing jobs — into repeatable workflows you can run from the terminal.

What problem does it solve?

Databricks is powerful, but many critical workflows remain:

  • UI-dependent
  • Stateful
  • Hard to reproduce
  • Hard to audit

Lakecode makes those workflows explicit and repeatable, producing evidence you can share.

Docs: How it works

Lakecode is built around a step-based workflow engine:

  • Deterministic steps: SQL execution, REST API calls, log fetches, metadata inspection
  • LLM steps: interpretation, summarization, prioritization, recommendations

Each workflow can:

  • Pass data between steps
  • Report progress
  • Retry on transient errors
  • Abort cleanly

Lakecode also maintains a lightweight workspace profile to reduce "where is that job/table/schema?" context switching.

Docs: Evidence packs

What is an evidence pack?

An evidence pack is a set of outputs (metrics, schema snapshots, samples, summaries) produced by a workflow so the result can be shared and reviewed.

Today

/prove produces a table health evidence pack.

Coming soon

Lakecode will standardize evidence packs across every workflow (debug, cost, governance, deploy), including structured JSON outputs and replay plans.

Docs: Configuration

Conventions

Lakecode loads conventions from:

  • ~/.lakecode/conventions.md
  • .lakecode/conventions.md (project-level)

Use conventions to teach lakecode:

  • Default catalog/schema
  • Naming patterns
  • Preferred warehouses
  • Team standards

Skills injection

Lakecode can load relevant Databricks skills based on keywords in your requests. Skills are treated as guidance; workflows remain deterministic where possible.

Docs: Safety & plan mode

Lakecode is designed to be safe by default:

  • Plan/approval flow for complex tasks
  • Guardrails for common mistakes
  • Output truncation to avoid flooding prompts and logs

As more write-capable workflows are added, policy packs and org-managed settings will become first-class.

Make Databricks work scriptable, explicit, and auditable.

Stop chasing UI workflows. Start producing evidence.